See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/337535111

Models and Algorithms for the Traveling Salesman Problem with Time-
Dependent Service Times

Article  in  European Journal of Operational Research · June 2020

DOI: 10.1016/j.ejor.2019.11.046

CITATION READS
1 170

3 authors:

Valentina Cacchiani Carlos Contreras Bolton

University of Bologna University of Concepción
49 PUBLICATIONS   1,519 CITATIONS    19 PUBLICATIONS   89 CITATIONS   

SEE PROFILE SEE PROFILE

Paolo Toth
University of Bologna
236 PUBLICATIONS   19,348 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Automatic Design of Algorithms View project

All content following this page was uploaded by Carlos Contreras Bolton on 01 December 2019.

The user has requested enhancement of the downloaded file.

 Models and Algorithms for the Traveling Salesman
Problem with Time-Dependent Service Times
Valentina Cacchiani, Carlos Contreras-Bolton, Paolo Toth
DEI, University of Bologna, Viale Risorgimento 2, I-40136 Bologna, Italy
{valentina.cacchiani, carlos.contrerasbolton, paolo.toth}@unibo.it

Abstract
The Traveling Salesman Problem with Time-dependent Service times (TSP-TS) is a general-
ization of the Asymmetric TSP, in which the service time at each customer is given by a (linear
or quadratic) function of the corresponding start time of service. TSP-TS calls for determining
a Hamiltonian tour (i.e. a tour visiting each customer exactly once) that minimizes the total
tour duration, given by the sum of travel and service times. We propose a new Mixed Integer
Programming model for TSP-TS, that is enhanced by lower and upper bounds that improve
previous bounds from the literature, and by incorporating exponentially many subtour elimina-
tion constraints, that are separated in a dynamic way. In addition, we develop a multi-operator
genetic algorithm and two Branch-and-Cut methods, based on the proposed model. The algo-
rithms are tested on benchmark symmetric instances from the literature, and compared with an
existing approach. The computational results show that the proposed exact methods are able
to prove the optimality of the solutions found for a larger set of instances in shorter computing
times. We also tested the Branch-and-Cut algorithms on larger size symmetric instances with
up to 58 nodes and on asymmetric instances with up to 45 nodes, demonstrating the effective-
ness of the proposed algorithms. In addition, we tested the genetic algorithm on symmetric and
asymmetric instances with up to 200 nodes.

Keywords: Travelling Salesman; Time-dependent Service times; Mathematical model; Branch-

and-Cut Algorithm; Genetic Algorithm

1 Introduction
The Traveling Salesman Problem with Time-dependent Service times (TSP-TS) has been recently
introduced in Taş et al. (2016), where Mixed Integer Programming (MIP) models and lower and
upper bounds were proposed. In the TSP-TS, the time to serve a customer is not assumed to be
constant, but, as it happens in some real-life applications, is defined by a function of the start time
of service at the customer. Before the work by Taş et al. (2016), service times were considered as
fixed values and directly included in the travel times. However, in practice, service times are not
always constant: for example, the availability of parking lots can be different at different times of
the day, or some areas can be limited to traffic in certain time periods. Therefore, the vehicle can
get closer to or farther from the customers to be visited, thus requiring shorter or longer service
times. Another example is the personnel availability for loading or unloading goods at customers,
that can vary over time. Similarly, the amount of goods in the customer warehouses can be smaller
at the beginning of the day and increase later on, thus making the service times needed to store
the goods at the customers variable during the day.
In the TSP literature, most of the works that take time dependency into account consider the
variability of the travel times, i.e., study the so-called Time-Dependent TSP (TDTSP). In several
papers, the arc cost is dependent on the position of the arc in the TSP tour. This problem is related
to single machine scheduling, in which a set of jobs needs to be scheduled, and the transition cost
cijt depends on the two consecutive jobs i and j and on the position t at which job i is processed. In

1
addition, set-up and cooling costs are considered, respectively, for the initial and final states of the
machine. The goal is to determine the minimum cost sequence of jobs. In Picard and Queyranne
(1978), three Integer Programming formulations are proposed: the first one is a generalization
of a TSP formulation in which costs are time-dependent, the second one refers to the Quadratic
Assignment Problem (QAP) and the third one is based on a layered graph. In the latter formulation,
the set of nodes of the layered graph contains the two copies of the depot 0 and n + 1, and nodes
(i, t) for each node i and position t. As observed by one of the reviewers of our paper, under the
hypothesis that all customers have the same service time function, TSP-TS can be formulated as a
TDTSP on the layered graph proposed in Picard and Queyranne (1978). We show, in the Appendix
(available as Supplementary Material online), how the layered graph formulation can be used to
model TSP-TS. A Branch-and-Bound algorithm based on a subgradient optimization procedure is
developed. In addition, the objective of minimizing the tardiness costs is considered. In Vander Wiel
and Sahinidis (1996), a Benders decomposition approach is applied to a MIP model, which is a
linearization of the quadratic formulation presented in Picard and Queyranne (1978), and combined
with Pareto-optimal Benders cuts. In Bigras et al. (2008), the two linear formulations proposed
in Picard and Queyranne (1978) are considered, and strengthened by applying Dantzig–Wolfe
decomposition and cuts for the classical TSP and the node packing problem. A Branch-and-
Cut-and-Price algorithm, based on these formulations, is developed. In Abeledo et al. (2013),
the polytope corresponding to the layered-graph based TDTSP linear formulation proposed in
Picard and Queyranne (1978) is studied, and facets are identified. A Branch-and-Cut-and-Price
algorithm, that includes the proposed cuts, is developed. In Miranda-Bront et al. (2014), two of
the formulations presented in Picard and Queyranne (1978) and Vander Wiel and Sahinidis (1996)
are studied and compared. Additional valid inequalities and facets are proposed, and embedded in
a Branch-and-Cut approach. Recently, Kinable et al. (2017) proposed a method for solving time-
dependent sequencing problems, i.e., sequencing problems in which setup times between two jobs
depend on the position of the jobs in the ordering. These problems include the TDTSP, the TDTSP
with time windows and the time-dependent sequential ordering problem. A hybrid method that
combines a constraint programming model with two relaxations is proposed. We refer the reader
to Gouveia and Voß (1995) for a classification of the formulations of the TDTSP.
The time dependency related to the travel times is also taken into account by another group
of works that, instead of considering that the arc costs depend on the position of the arc in the
tour, consider that the cost (or travel time) τij (t) of arc (i, j) depends on the time t at which the
vehicle leaves node i. In Cordeau et al. (2012), the time horizon is partitioned into intervals, and
the average speed value in each interval is assumed to be known. The travel time τij (t) is then
computed according to the function proposed in Ichoua et al. (2003), that considers the speed to be
constant during an interval, and allows it to change when moving to the following interval, i.e., the
travel speeds are constant piecewise. The travel speed along an arc and during an interval is defined
in Cordeau et al. (2012) as dependent on the maximum arc travel speed, the lightest congestion
factor (i.e., the best congestion during the interval) and the degradation of the congestion factor
w.r.t the least congested arc. In Cordeau et al. (2012), an Integer Linear Programming (ILP) model,
including subtour elimination constraints, is proposed for this problem, and valid inequalities are
derived. A Branch-and-Cut algorithm is developed and tested on a set of instances having different
traffic patterns. In Arigliano et al. (2018), a Branch-and-Bound algorithm is proposed for the same
problem, and tested on the same instances and on larger size ones (with up to 50 customers). The
reported computational results show that this algorithm outperforms the approach proposed in
Cordeau et al. (2012).
Finally, more recently, the problem studied in Cordeau et al. (2012) has been extended to include
time window constraints. This extension was first proposed in Arigliano et al. (2019), where an

2
ILP model and valid inequalities were proposed. In addition, a Branch-and-Cut algorithm was
developed and tested on instances adapted from those considered in Cordeau et al. (2012). In
Montero et al. (2017), an alternative formulation, based on the model proposed in Sun et al.
(2018) for the Profitable Time-Dependent TSP with Time Windows and Pickup and Delivery, was
presented, and a Branch-and-Cut algorithm was developed. The computational results show that
this algorithm obtains many additional optimal solutions compared to the approach presented in
Arigliano et al. (2019). The recent work Vu et al. (2018) also studies the time-dependent TSP with
time windows. A solution method, based on the representation of the problem on a time-expanded
network, is proposed. The method relies on the Dynamic Discretization Discovery framework,
proposed in Boland et al. (2017), that works on a partially time-expanded network, and iteratively
refines it until optimality is guaranteed. This new approach is also tested on the benchmark
instances considered in Arigliano et al. (2019) and obtains the optimal solution for all the instances
in shorter computing times.

1.1 Contribution
To the best of our knowledge, the only work studying the TSP-TS is Taş et al. (2016), where
compact Mixed Integer Programming (MIP) models, i.e. formulations having a polynomial number
of variables and constraints, have been proposed, together with lower and upper bounds aimed at
improving the solution process. In this paper, we propose a new MIP model for the TSP-TS, that
can be used with linear or quadratic service time functions, and that embeds novel improved lower
and upper bounds. In this model, we consider exponentially many subtour elimination constraints,
that are separated dynamically. In addition, we developed a multi-operator Genetic Algorithm
(GA), that includes many crossover and mutation operators from the literature, and inserts, in
the initial population, solutions generated by using the solution of the continuous relaxation of the
MIP model. The main contribution of this work consists of the design and development of two
Branch-and-Cut (B&C) algorithms. Both of them are based on the presented model. In particular,
we propose a B&C algorithm that includes all the improved bounds and the subtour elimination
constraints separation. In addition, we propose a Dynamic B&C algorithm, in which the lower
and upper bounds are dynamically updated at each node of the decision tree. The GA, B&C and
Dynamic B&C algorithms are tested on symmetric benchmark instances from Taş et al. (2016),
and on asymmetric instances with up to 45 nodes and larger size symmetric instances with up to
58 nodes from the TSPLIB library (Reinelt (1991)). Finally, the GA is tested on symmetric and
asymmetric instances from the TSPLIB library with up to 200 nodes.
Section 2 presents the problem definition and introduces the used notation. In Section 3 we
report the mathematical models and the lower and upper bounds proposed in Taş et al. (2016).
Section 4 describes the new mathematical model and the improvements to the lower and upper
bounds from the literature. Section 5 is devoted to the description of the solution methods, i.e. the
GA, B&C and Dynamic B&C algorithms, and in Section 6 we report the obtained results and the
comparison with the method presented in Taş et al. (2016). Finally, we conclude our paper with
some remarks in Section 7.

2 Problem Definition
TSP-TS is defined on a complete directed graph G = (N, A). The set of nodes N = {0, 1, . . . , n, n +
1} contains the set of customers N \ {0, n + 1} and the depot represented by nodes 0 and n + 1
(the depot is duplicated for convenience). The set of arcs A contains one arc (i, j) for each pair of
nodes and has an associated travel time tij , as in the classical Asymmetric TSP (ATSP) (Öncan

3
et al. (2009), Roberti and Toth (2012)). In addition, each customer i ∈ N \ {0, n + 1} requires a
service time, defined as a continuous function si (bi ), where bi represents the start time of service
at node i. We set b0 , s0 (b0 ), and sn+1 (bn+1 ) to 0, since the depot does not require any delivery of
goods.
In TSP-TS, we have to determine the Hamiltonian path (i.e., the path visiting each node
exactly once) from node 0 to node n + 1 that minimizes the total path duration, given by the sum
of the total travel and service times. In the following, we will use the terms “path” or “route” as
synonyms, since nodes 0 and n + 1 are two copies of the same depot. Since TSP-TS generalizes
ATSP, that occurs when si (bi ) = 0 (i ∈ N \ {0, n + 1}), it is NP-hard. In addition, as observed in
Taş et al. (2016), an optimal ATSP solution is not always optimal for TSP-TS. We next report three
important properties proven in Taş et al. (2016) that are used for the definition of the mathematical
models and of the lower and upper bounds.
Property 1. First-In-First-Out property (FIFO): If the service at customer i starts at a time bi ,
any service starting at a later time at that customer cannot be completed earlier than bi + si (bi ).
In addition, in Taş et al. (2016), the authors proved that si (bi ) satisfies the FIFO property if and
only if dsdb
i (bi )
i
≥ −1.
0
Property 2. Waiting property: If the vehicle arrives at customer i before bi (the earliest time at
which the FIFO property starts holding at that customer), waiting at that customer to begin service
0
at time bi is then beneficial.
Property 3. Arrival time property: If all customers have the same service time function, then the
waiting time required to satisfy the FIFO property can be spent at the depot.
As in Taş et al. (2016), we will consider service time functions that satisfy the FIFO property
for any value of the arrival time, and we will assume that the arrival time property holds, i.e., the
start time of service at a customer coincides with the arrival time at that customer.

3 Mathematical Models and Bounds from the Literature

In this section, we report the MIP models and the lower and upper bounds proposed in Taş et al.
(2016). The best method proposed in that paper will be used for comparison in Section 6.

3.1 Basic Model

We present the basic model that is a natural formulation for the TSP-TS. For every arc (i, j) ∈ A,
we introduce a binary variable xij (i ∈ N, j ∈ N ) assuming value 1 if node j is served immediately
after node i (and 0 otherwise). In addition, for every node i ∈ N , we introduce a non-negative
variable bi representing the arrival time at node i. According to the arrival time property, bi also
corresponds to the start time of service at i. The basic model reads as follows:

XX X
min tij xij + si (bi ) (1)
i∈N j∈N i∈N \{0,n+1}
X
xij = 1, i ∈ N \ {n + 1}, (2)
j∈N \{i}
X
xij = 1, j ∈ N \ {0}, (3)
i∈N \{j}

4
 bi + si (bi ) + tij − M (1 − xij ) ≤ bj , i ∈ N, j ∈ N, (4)
bi ≥ 0, i ∈ N, (5)
xij ∈ {0, 1}, i ∈ N, j ∈ N. (6)

The objective function (1) minimizes the total duration of the Hamiltonian path from node 0 to
node n + 1, where the duration is given by the sum of the total travel times (first summation) and
of the total service times (second summation). As in the classical ATSP, we impose, by constraints
(2) and (3) respectively, to have exactly one outgoing arc from each node except for the depot n+1,
and one ingoing arc for each node, except for the depot 0. The novel constraints (4) are specific for
the TSP-TS and require that, if node j is visited immediately after node i, the start time of service
bj at node j must be at least the start time of service bi at node i plus the service time si (bi ) at
node i plus the travel time from i to j. To deactivate these constraints when the corresponding
arc (i, j) is not selected, a large positive constant M is used, whose value is determined by a lower
bounding procedure described in the following. Note that these constraints guarantee subtours
to be infeasible, thus making model (1)-(6) valid without explicit subtour elimination constraints.
Finally, (5) and (6) define the domain of the variables.
The lower bounding procedure proposed in Taş et al. (2016) for determining the value of M is
proven to be valid when all customers have the same (quadratic or linear) service time function,
and works as follows. The first n + 1 largest travel times are selected and taken in descending order.
Let L(1), . . . , L(n + 1) be the ordered list of the selected travel times. Then, the “path” consisting
of the sequence of these travel times is considered, and the arrival (bi ) and departure (di ) times at
each node i in the path are computed as follows:

• b1 = L(1); di = bi + si (bi ), i = 1, . . . , n

• bi = di−1 + L(i), i = 2, . . . , n + 1.

The M value is set equal to bn+1 , i.e., it corresponds to the arrival time at the depot. This value
is sufficient to deactivate constraints (4) for arcs (i, j) ∈ A having xij = 0.

3.2 Gavish and Graves Model

In Taş et al. (2016), in order to strengthen model (1)-(6), the polynomially many Gavish and Graves
(GG) subtour elimination constraints (proposed in Gavish and Graves (1978) for the ATSP) were
added. Let gij (i ∈ N \ {0, n + 1}, j ∈ N \ {0}) be a non-negative variable denoting the number of
arcs in a path from depot 0 to arc (i, j) ∈ A. The GG constraints read as follows:

X X
gij − gji = 1, i = 1, . . . , n, (7)
j∈N \{0} j∈N \{0,n+1}

0 ≤ gij ≤ nxij , i = 1, . . . , n, j = 1, . . . , n + 1. (8)

Constraints (7) require that the number of arcs in a path from depot 0 to node i (whatever
successor node is selected from i), must be one unit larger than the number of arcs in a path from
depot 0 to the predecessor of node i (whatever node it is). Constraints (8) require the gij variables
to be non-negative and restrict their value to be 0, if arc (i, j) is not selected, and at most n, if it
is selected, since the length of the full Hamiltonian path is n + 1.

5
3.3 Lower and Upper Bounds
In Taş et al. (2016), three bounds were proposed:

• upper bound on the total route duration (B1 ),

• lower bound on the total service time (B2 ),

• lower and upper bounds on the arrival time at each customer (B3 ).

The upper bound B1 on the total route duration of the optimal solution is presented for
the case in which all customers have the same service time function. It consists of applying
the Nearest Neighbor Heuristic (NNH) algorithm for the ATSP. The obtained sequence of nodes
(0, a(1), a(2), . . . , a(n), a(n+1) = n+1) is used to define the list of travel times L(1), L(2), . . . , L(n+
1): in particular, L(1) = t0,a(1) and L(k) = ta(k−1),a(k) (k = 2, . . . , n + 1). Then, arrival (bi ) and
departure (di ) times at each node i in the path are computed, as described for the computation of
M . The value of the upper bound B1 is set equal to the arrival time bn+1 at the depot.
The lower bound B2 on the total service time is proposed for the case in which all customers
have the same linear service time function si (bi ) = βbi + γ, where β, γ > 0, and si (bi ) > 0
(i ∈ N \ {0, n + 1}). The value of B2 is computed by considering the first n smallest travel times.
More precisely, let L(1), L(2), . . . , L(n) be the ascending ordered list of the first n smallest travel
times, consider the path (0, 1, 2, . . . , n) whose arcs have the selected travel times, and compute the
arrival (bi ) and departure (diP ) times at each node i in the path, as for the computation of M . The
value of B2 is given by dn − ni=1 L(i), i.e., it is obtained by considering the departure time from
the last visited customer n and subtracting the total path travel time. Since the linear service
time function is non-decreasing, arriving at a node earlier gives a smaller service time than arriving
there later. As a consequence, a valid lower bound on the total service time can be obtained by
considering the path defined by the n smallest travel times. A more formal proof of the validity of
B2 can be found in Taş et al. (2016).
Finally, the upper and lower bounds B3 on the arrival time at each customer i ∈ N \ {0, n + 1}
are set, respectively, as: bi ≤ M and bi ≥ LB3 (i) = earliest time at which the FIFO property starts
holding for customer i, which is 0 for the considered benchmark instances.
The best method proposed in Taş et al. (2016) consists of the basic model (1)-(6) (where the
value of M is computed as described in Section 3.1), with the additional GG (polynomially many)
subtour elimination constraints (7)-(8) and the lower and upper bounds B1 , B2 and B3 . We identify
this method as TGJL16.

4 Improvements to Models and Bounds

In Section 4.1, we propose improved lower and upper bounds, and an improved value for M . In
addition, in Section 4.2, we enhance the basic model by using exponentially many explicit subtour
elimination constraints (SECs). Furthermore, in Section 4.3, we propose a new model for the
TSP-TS, which can also embed the improved bounds and the SECs.

4.1 Improved Bounds (IBs) and M

As described in Section 3, in Taş et al. (2016), three bounds were proposed to strengthen the basic
model. We improve these three bounds, under the same assumptions, by proposing:

• an improved upper bound on the total route duration (IB1 ),

6
 • an improved lower bound on the total service time (IB2 ),

• improved lower and upper bounds on the arrival time at each customer (IB3 ).

To compute an upper bound IB1 on the total route duration, instead of applying the NNH
algorithm, we developed a multi-operator Genetic Algorithm (GA), that provides high quality
solutions for the TSP-TS. Algorithm GA is based on the calibration of the probabilities to be
used for executing the considered crossover and mutation operators, performed according to the
algorithm proposed in Contreras-Bolton and Parada (2015) for the TSP, but it also relies on the
solution of the continuous relaxation of the basic model. In particular, as it will be shown in the
computational results (Section 6.1), GA obtains better solutions when the SECs are included in
the model. Algorithm GA will be described in Section 5.1. It provides the sequence of nodes of the
Hamiltonian path, that is used to compute the arrival and departure times at each node, and IB1
is set equal to bn+1 .
The improved lower bound on the total service time IB2 is proposed under the same assumptions
considered for B1 , i.e., all customers have the same linear service time function si (bi ) = βbi + γ,
where β, γ > 0, and si (bi ) > 0 (i ∈ N \ {0, n + 1}). The value of IB2 is computed as the maximum
of two lower bounds IBL2 and IBE2 , obtained as follows. In IBL2 , instead of considering the
first n smallest travel times as done in Taş et al. (2016), we proceed as follows. For each node
i (i = 0, 1, . . . , n), we compute the minimum and second minimum travel time arcs (i, mini ) and
(i, smini ), respectively, leaving node i, and set: τi = ti,mini = min{tij : j = 1, . . . , n, j 6= i} and
τi0 = ti,smini = min{tij : j = 1, . . . , n, j 6= i, j 6= mini }. Then, we execute the following four steps:

1. sort the n minimum travel time arcs (i, mini ) (i = 1, . . . , n) in ascending order according to
the travel times τi . Let ordh (h = 1, . . . , n) be the initial node of the hth ordered arc.

2. for each node i (i = 1, . . . , n) set fi = 0;

For h = 1, . . . , n:

(a) set i = ordh , j = mini ;

(b) if fj = 1 and minj = i and τj0 > τi then set L(h + 1) = min{τj0 , τi0 }
(c) else set L(h + 1) = τi , fi = 1.

3. order the n travel times L(h) (h = 2, . . . , n + 1) according to ascending values.

4. set L(1) = τ0 , consider the path (0, 1, . . . , n) whose arcs have the travel times L(1), L(2), . . . , L(n),
and proceed as in the computation of B2 , P i.e., compute the arrival (bi ) and departure (di )
times at each node i and set IBL2 = dn − ni=1 L(i).

At step 2, for each node i (i = 1, . . . , n), fi is equal to 1 if the value τi , corresponding to the travel
time of arc (i, mini ), has been assigned to L(h + 1), with i = ordh , i.e., if arc (i, mini ) is used in the
computation of IBL2 ; in this case, the reverse arc (mini , i) is not used in the following iterations
of the loop. Otherwise, the value fi keeps its initial value 0; in this case the reverse arc (mini , i)
can possibly be used.
For IBE2 , in a similar way, we compute, for each node i (i = 1, 2, . . . , n), the minimum and
second minimum travel time arcs (mini , i) and (smini , i), respectively, entering node i, and set
τi = tmini ,i = min{tji : j = 0, . . . , n, j 6= i} and τi0 = tsmini ,i = min{tji : j = 0, . . . , n, j 6= i, j 6=
mini }. Then we execute the same four steps described above for the computation of IBL2 , where,
in step 1, we consider arcs (mini , i) in place of arcs (i, mini ) (i = 1, . . . , n), and ordh (h = 1, . . . , n)
as the final node of the hth ordered arc, and, in step 4, we consider the path (0, 1, . . . , n) whose

7
arcs have the travel times L(2), L(3), . . . , L(n + 1). The improved lower bound IB2 is then set as
max{IBL2 , IBE2 }. The validity of IB2 relies on the fact that, in any Hamiltonian path, every node
(but the depot) must be visited exactly once. Therefore, we need to reach each node by selecting
only one of its leaving (entering, resp.) arcs. Since, as previously mentioned, the linear service
time function is non-decreasing, arriving at a node earlier gives a smaller service time than arriving
there later. For this reason, we can consider the smallest travel time arc leaving (entering, resp.)
every node. In particular, in the computation of IBL2 , we need to leave each node i (i = 1, . . . , n),
but the last node of the Hamiltonian path, by selecting only one of its leaving arcs (i, j), with
j = 1, . . . , n. In addition, the first arc of the Hamiltonian path must leave the initial depot 0: so,
as first arc of the path, we select the minimum travel time arc leaving the depot. For the last node
of the path, we consider the node say l, for which the associated τl has the largest value, given by
L(n + 1). In the computation of IBE2 , we need to enter each node i (i = 1, . . . , n) by selecting
only one of its entering arcs (j, i), with j = 0, 1, . . . , n.
Improved lower and upper bounds IB3 on the service start time at each customer i (i =
1, 2, . . . , n) can be computed by considering that every node must be reached starting from the
depot 0 and must reach depot n + 1. Therefore, for every customer i, the start time of the service
(which corresponds to the arrival time at the customer thanks to the FIFO property) cannot be
smaller than the time corresponding to the shortest path from the depot 0 to i. In other words,
bi ≥ sp0,i , where sp0,i is the value of the shortest path from depot 0 to customer i, which also
includes in the computation the service times at the visited nodes, except for the service time at i.
In addition, in order to determine a solution not worse than that obtained by the GA algorithm,
the start time of service at each customer i cannot be larger than the total route duration found
by GA, i.e., bi ≤ IB1 . The latter upper bound can be further improved, by considering that the
vehicle must go back to the depot, i.e., the depot n + 1 must be reached from every customer i.
This gives the following upper bound on the service start time of customer i: bi ≤ IB1 − spi,n+1 ,
where spi,n+1 gives the value of the shortest path from i to n + 1, by taking into account the service
times at the visited nodes, including the service time at i. Note that, for the quadratic service time
function, in the computation of the shortest path spi,n+1 we only consider the travel times.
Finally, when all customers have the same (quadratic or linear) service time function, as assumed
in Taş et al. (2016), we propose an improvement of the value of M . Recall that, in constraints
(4), the value of M is used to deactivate the constraint, when arc (i, j) is not selected. The value
of M must be such that bj ≥ −M + bi + si (bi ) + tij , when arc (i, j) is not in the solution, i.e.,
M ≥ bi + si (bi ) + tij − bj . We consider the value IB1 of the feasible solution obtained by the GA
algorithm, which represents the best available value of the total route duration, i.e., the arrival
time at node (n + 1). Of course, any upper bound on the total route duration could be used
instead of IB1 . However, this value is not necessarily enough to deactivate constraints (4), since
arc (i, j) might not be selected in the GA solution and might be an arc with a “long” travel time tij .
Therefore, the value of M is chosen as dependent on the specific arc (i, j), i.e., we use Mij instead
of M . We set M0j = t0j for j = 1, . . . , n, and Mij = IB1 + tij for i = 1, . . . , n, j = 1, . . . , n + 1.
Note that replacing M by Mij for each arc (i, j) ∈ A is correct by considering that:

• for arcs (0, j) (j = 1, . . . , n), M0j = t0j ≥ b0 + s0 (b0 ) + t0j − bj , since b0 = s0 (b0 ) = 0;

• for arcs (i, j) (i = 1, . . . , n, j = 1, . . . , n + 1), Mij = IB1 + tij ≥ bi + si (bi ) + tij − bj , since
bi + si (bi ) ≤ IB1 (as IB1 is an upper bound on the total route duration).

8
4.2 Subtour Elimination Constraints (SECs)
In Taş et al. (2016), additional subtour elimination constraints, namely the GG constraints, are
used to improve the basic model performance. The authors also evaluate two other compact ATSP
formulations with a polynomial number of subtour elimination constraints, obtained, respectively,
by using the Miller, Tucker and Zemlin (MTZ) (Miller et al. (1960)) constraints and the Desrochers
and Laporte (DL) (Desrochers and Laporte (1991)) constraints to improve the Linear Program-
ming (LP) relaxation. These constraints are incorporated in the basic model as valid inequalities.
However, the best performance was achieved by considering the GG constraints. In this paper,
we propose to replace the GG constraints (7)-(8) with the explicit subtour elimination constraints
(SECs) proposed in Dantzig et al. (1954) for the ATSP:
X X
xij ≥ 1, S ⊆ N \ {n + 1}, 0 ∈ S, |S| ≥ 2. (9)
i∈S j∈N \S

These constraints, imposed for every subset S of nodes that contains at least two nodes, the depot
0, and does not contain depot n + 1, require to select at least one arc going from a node in S to
a node not in S. Since the number of constraints (9) is exponential in the number of nodes, we
impose in the model only a subset of these constraints, and iteratively add violated constraints
detected by applying the constraint separation procedure proposed in Padberg and Rinaldi (1990).

4.3 A New Model (NM) for the TSP-TS

The new model (NM) described in this section is based on the formulation proposed in Maffioli and
Sciomachen (1997) for the ATSP with Time Windows (ATSP-TW). For sake of clarity, we start by
presenting this model, and then we show how to modify it to include the service time component.
For every arc (i, j) ∈ A, let xij be a binary variable assuming value 1 if arc (i, j) is selected in
the optimal route (and 0 otherwise). For every customer i ∈ N , let [ri , di ] be the corresponding
time window. Moreover, for i ∈ N \ {0, n + 1} and j ∈ N \ {0}, i 6= j, let yij be an additional
variable with the following meaning:

• if xij = 0 then yij = 0;

• if xij = 1 then yij denotes the arrival time at node i when node j is visited immediately after
node i.

The ATSP-TW model presented in Maffioli and Sciomachen (1997) reads as follows:

XX
min tij xij (10)
i∈N j∈N
X
xij = 1, i ∈ N \ {n + 1}, (11)
j∈N \{i}
X
xij = 1, j ∈ N \ {0}, (12)
i∈N \{j}

ri xij ≤ yij ≤ di xij , i ∈ N \ {0, n + 1}, j ∈ N \ {0}, i 6= j, (13)

X X X
yij + tij xij ≤ yjk , j ∈ N \ {0, n + 1}, (14)
i∈N \{0,n+1},i6=j i∈N \{n+1},i6=j k∈N \{0},k6=j

xij ∈ {0, 1}, i ∈ N, j ∈ N, (15)

9
 yij ≥ 0, i ∈ N \ {0, n + 1}, j ∈ N \ {0}, i 6= j. (16)

The objective function (10) calls for minimizing the total travel times. Constraints (11) and
(12) require, respectively, to have an outgoing arc from every node except for the depot n + 1, and
an ingoing arc for every node except for the depot 0. Time window constraints are imposed by
(13), where the arrival time yij at node i (when j is visited immediately after i) is required to be
within the time window [ri , di ]. These constraints are also used to set yij to 0 when the arc (i, j)
is not selected in the optimal route. Constraints (14) are used to specify the arrival time yij at
every customer j ∈ N \ {0, n + 1}: when node j is visited immediately after node i, the arrival time
yjk at customer j, no matter which node k is visited afterwards (including depot n + 1), must be
at least the arrival time yij at node i, i.e., any predecessor of customer j (except for the depot),
plus the travel time tij from node i to node j. Finally, constraints (15) and (16) define the variable
domains.
The new proposed model (NM) uses, in addition to the xij and yij variables, the bi variables
defined in model (1)-(6), representing the start time of service at node i ∈ N . We first present
NM for the case of linear service time function si (bi ) = βbi + γ, where β, γ > 0, and si (bi ) > 0
(i ∈ N \ {0, n + 1}), and then extend it to a quadratic service time function. Model (10)-(16) is
modified to deal with linear time-dependent service times, as follows:

XX X
min tij xij + si (bi ) (17)
i∈N j∈N i∈N \{0,n+1}
X
xij = 1, i ∈ N \ {n + 1}, (18)
j∈N \{i}
X
xij = 1, j ∈ N \ {0}, (19)
i∈N \{j}
X
bi = yik , i ∈ N \ {0, n + 1}, (20)
k∈N \{0},k6=i

yij ≥ LB3 (i)xij , i ∈ N \ {0, n + 1}, j ∈ N \ {0}, i 6= j, (21)

yij + βyij + γxij ≤ M xij i ∈ N \ {0, n + 1}, j ∈ N \ {0}, i 6= j, (22)
X
yij +
i∈N \{0,n+1},i6=j
X X
βyij + γxij +
i∈N \{0,n+1},i6=j i∈N \{n+1},i6=j
X X
tij xij ≤ yjk , j ∈ N \ {0, n + 1}, (23)
i∈N \{n+1},i6=j k∈N \{0},k6=j

bi ≥ 0, i ∈ N, (24)
xij ∈ {0, 1}, i ∈ N, j ∈ N, (25)
yij ≥ 0, i ∈ N \ {0, n + 1}, j ∈ N \ {0}, i 6= j. (26)

The objective function (17) is the same as (1), and minimizes the total route duration given
by the sum of the total travel and service times. As in the previous models, constraints (18) and
(19) impose to select an outgoing arc from every node except for the depot n + 1 and an ingoing
arc for every node except for the depot 0. Constraints (20) are used to define the values of the bi
variables: the arrival time bi at i can be expressed as the arrival time yik , no matter which successor

10
node k is chosen. Constraints (21) and (22), in which lower and upper bounds on the arrival times
at the customers are used, replace the time window constraints (13). In particular, we use the
bounds B3 proposed in Taş et al. (2016). Note that constraints (21) and (22) ensure that yij is
set to 0 if xij is also 0. Constraints (23) correspond to (14) but also include the service time at
the customers: for every customer j, when node j is visited immediately after node i, the arrival
time yjk at j (no matter which successor k is selected) must be at least the arrival time yij at its
predecessor i (no matter which it is) plus the service time at i (given by βbi + γ), plus the travel
time tij between i and j. Constraints (24)-(26) restrict the variable domain. It can be observed
that the constraints (23) can be written as equality constraints when the arrival time property (see
Property 3 in Section 2) holds. We performed computational experiments, on the instances with
the linear service time functions proposed in Taş et al. (2016), by replacing the constraints (23)
with the corresponding equality constraints. Similar performances were obtained by the proposed
Dynamic B&C algorithm (described in Section 5.2) in both cases: namely, the same number of
instances was solved to optimality within the given time limit, and the average optimality gap was
slightly larger when solving the model with the equality constraints than with the constraints (23).
Therefore, we do not report the computational results obtained by the Dynamic B&C algorithm
with the equality constraints.
In order to improve NM, we can insert the improved bounds IB3 in constraints (21) and (22)
as follows:

yij ≥ sp0,i xij , i ∈ N \ {0, n + 1}, j ∈ N \ {0}, i 6= j, (27)

yij + βyij + γxij ≤ (IB1 − spi,n+1 )xij i ∈ N \ {0, n + 1}, j ∈ N \ {0}, i 6= j. (28)

In particular, if arc (i, j) is selected, the arrival time yij at node i, when j is its successor, must
be at least the value of the shortest path from the depot 0 to i. In addition, when arc (i, j) is
selected, the arrival time yij at node i, when j is its successor, plus the service time si (bi ), which
corresponds to βbi + γ, must be at most the improved upper bound IB1 on the total route duration
minus the value of the shortest path from i to the depot n + 1.
If we consider a quadratic service time function defined as si (bi ) = αb2i − βbi + γ, as done in
Taş et al. (2016), constraints (22) and (23) are changed as follows.
2
yij + αyij − βyij + γxij ≤ M xij i ∈ N \ {0, n + 1}, j ∈ N \ {0}, i 6= j, (29)
X
yij +
i∈N \{0,n+1},i6=j
X X X
2
αyij − βyij + γxij +
i∈N \{0,n+1},i6=j i∈N \{0,n+1},i6=j i∈N \{n+1},i6=j
X X
tij xij ≤ yjk , j ∈ N \ {0, n + 1}. (30)
i∈N \{n+1},i6=j k∈N \{0},k6=j

A similar improvement as in the case of the linear service time functions can be applied for the
quadratic case by replacing constraints (29) with the following ones:
2
yij + αyij − βyij + γxij ≤ (IB1 − spi,n+1 )xij i ∈ N \ {0, n + 1}, j ∈ N \ {0}, i 6= j. (31)

In the following, we will call NMB the model corresponding to (17)-(26), for the linear service
time functions, and to (17)-(21), (24)-(26), (29)-(30), for the quadratic service time function. We
will call NMI the improved model corresponding to (17)-(20), (23)-(26), (27)-(28), for the linear
service time functions, and to (17)-(20), (24)-(26), (27), (30)-(31), for the quadratic service time
function.

11
5 Solution Methods
To solve the TSP-TS we propose the metaheuristic algorithm GA, presented in Section 5.1, that
provides high quality solutions in very short computing times (see Section 6.1), and two exact B&C
algorithms, described in Section 5.2, that use all the improved bounds (for the quadratic service
times, IB2 is not used), the improved value of M and the SECs separation. The B&C algorithms
are applied using the new model NMI and, for comparison, the basic model.

5.1 Genetic Algorithm

Genetic algorithms are effective metaheuristic algorithms, that belong to the class of evolutionary
algorithms, and are based on the evolution of a population of individuals, corresponding to solu-
tions of the considered problem. In the proposed GA algorithm, each individual corresponds to a
Hamiltonian path from depot 0 to depot n + 1, and we represent it as a permutation of nodes in
{1, . . . , n}.
An initial population is computed by considering only the nodes in N \ {n + 1} and by applying
three algorithms, each one generating a given percentage of the population:

• an algorithm (random route algorithm) that generates a random route (25% of the initial
population): in particular, starting from the depot 0, we randomly choose a node, and then,
from the chosen node, we select the next node in a random way among the non-visited ones,
and so on, until a Hamiltonian path is built;

• the randomized NNH algorithm with each individual generated starting from a different
randomly chosen node (not necessarily the depot), and finishing when all the nodes in N \
{n + 1} have been visited (50% of the initial population);

• a continuous relaxation based algorithm, that uses the optimal continuous relaxation solution
x of the basic model with the improved bound IB2 to build a feasible Hamiltonian path (25%
of the initial population).

We use the basic model, instead of model NMI, in order to reduce the global computing time
of algorithm GA. In particular, the first individual of the population is obtained in a deterministic
way as follows: we define the depot 0 as the starting node h, and iteratively select the non-visited
node j such that xhj + xjh has the highest value; arc (h, j) is added to the route, and j becomes
the new starting node h. If, for all the non-visited nodes j we have xhj + xjh = 0, then we choose
the arc with the smallest thj . The procedure is repeated until a complete feasible Hamiltonian
path has been obtained. Then, the remaining individuals are obtained by adding randomness in
the construction of the route: we start from the depot 0, defined as node h, and select the next
non-visited node j according to a probability given by xhj + xjh . If, for all the non-visited nodes j,
we have xhj + xjh = 0, then we choose the arc with the smallest thj . The selected node j becomes
the new starting node h, and the procedure is repeated until a Hamiltonian path has been built.
We propose two variants for the continuous relaxation based algorithm, which either consider or
not the SECs separation during the solution of the relaxed model. A comparison of the two variants
is reported in Section 6.1.
We consider many crossover and mutation operators, listed in the following. We omit their
description and refer the interested reader to the corresponding reference: Order Based Crossover
(OX2) (Syswerda (1991)), Distance Preserving Crossover (DPX)(Reisleben et al. (1996)), Maximal
Preservative Crossover (MPX) (Mühlenbein (1991)), Heuristic Crossover (HX) (Grefenstette et al.
(1985)), Modified Inver-over Operator (MIO) (Wang et al. (2012)), Uniform Nearest Neighbor

12
(UNN) (Buriol et al. (2004)), Exchange Mutation (EM) (Banzhaf (1990)), Scramble mutation
(SM) (Syswerda (1991)), Inversion mutation (IVM) (Fogel (1993)), Insertion Mutation (ISM) (Fogel
(1988)), Greedy Sub-Tour Mutation (GSTM) (Albayrak and Allahverdi (2011)) and 3-opt (we use
a simplified 3-opt version, which considers all pairs of arcs and selects the third arc randomly out
of ten arcs).
We performed a calibration of the probabilities to use for each crossover and mutation operator,
by executing the algorithm proposed in Contreras-Bolton and Parada (2015), which considers many
alternative operators and finds a good combination of them. In that paper, the authors studied
how to select appropriate combinations of crossover and mutation operators, typically used for
the TSP, and applied evolutionary computing to determine the best probability for each operator.
The algorithm proposed in Contreras-Bolton and Parada (2015) was calibrated on a subset of the
TSP-TS instances considered in Section 6 (three symmetric ones and three asymmetric ones) and
by considering the small linear service time function si = 5(10−3 )bi + 3(10−2 ) (i ∈ N \ {0, n + 1}).
Then, the same probabilities are used in all the computational experiments.
We consider an initial population of 150 individuals, that can contain duplicated individuals.
Each individual is evaluated (by starting from depot 0), according to the objective function (1), that
represents the total route duration. An iterative loop is then executed to evolve the population, until
a terminating condition is met (in our computational experiments, 200 iterations are performed).
At each iteration, the following steps are executed:

1. Selection is used to select an individual in the population. It is a tournament selection based

on three individuals. In particular, three individuals are randomly chosen from the current
population and the best one is selected.

2. Crossover is used to combine two individuals of the population. When the two individuals
have been selected (according to the tournament selection), one of the crossover operators,
according to the probabilities reported in Table 1, is applied. The crossover operator is applied
on the two individuals (parents) with 85% probability.

3. Mutation is used to modify an individual in order to add diversity to the population and to
ensure the exploration of a large solution space. It is applied on the offspring generated at
step 2, with 20% probability. One of the mutation operators, according to the probabilities
shown in Table 1, is selected and applied to obtain a new individual.

4. Evaluation is applied to compute the quality of the obtained solutions: the offspring generated
by crossover and mutation are evaluated, according to the objective function (1).

5. Elitism is applied as a last step: the offspring population created by selection, crossover and
mutation replaces the original parental population, but the 10% worst offspring are replaced
by the best parents in the current population.

Despite its simplicity, GA is able to find, in very short computing times, solutions of the TSP-
TS with small percentage gaps with respect to the best known solution values, as shown in Section
6.1.

13
 Crossover Prob. Mutation Prob.
OX2 0.010 EM 0.100
DPX 0.400 SM 0.009
MPX 0.076 IVM 0.008
HX 0.214 ISM 0.018
MIO 0.100 GSTM 0.425
UNN 0.200 3-opt 0.440

Table 1: Probabilities of using each crossover or mutation operator.

5.2 Branch-and-Cut and Dynamic Branch-and-Cut Algorithms

We propose a B&C algorithm, in which the SECs are separated at every node of the decision
tree by using the separation procedure proposed in Padberg and Rinaldi (1990) until no violated
constraints exist. The B&C algorithm is based on the new model NMI, in which all the improved
bounds IB1 , IB2 and IB3 , as well as the improved value of M , are used. Note that, since IB2
is only valid for linear service time functions, as explained in Section 4.1, we do not apply it on
instances with a quadratic service time function. The GA algorithm is used to define the value
of IB1 , and for the upper bound on the start time of service in IB3 . In the B&C algorithm, we
apply the branching rules chosen by default in the CPLEX solver, and use the callback functions
to separate the SECs.
Beside the B&C algorithm, we also propose a Dynamic B&C algorithm, in which the improved
bounds IB2 and IB3 are dynamically updated during the solution process in order to take into
account the branching decisions. In particular, the two bounds are recomputed at every node α of
the decision tree, by taking into account that, for the currently considered node α, some arcs (i, j)
have been selected in the solution (xij = 1) and must be chosen in the bound computation, and
other arcs have been discarded (xij = 0) and may not be used in the computation of the bounds.
In this way, the bounds can be tightened using local information from the node. More precisely,
for the currently considered node α, let IA (resp. EA) be the set of arcs (i, j) ∈ A such that the
variable xij has been fixed to 1 (resp. to 0) at the previous levels of the decision tree. Then an
updated travel time matrix (t0ij ) is defined as follows: (i) for i ∈ N \ {n + 1} and j ∈ N \ {0}
set t0ij = tij ; (ii) for (i, j) ∈ EA set t0ij = ∞; (iii) for (i, j) ∈ IA set t0ik = ∞ for k ∈ N \ {0, j}
and t0kj = ∞ for k ∈ N \ {n + 1, i}. The updated improved bounds IB2 and IB3 at node α are
computed by considering matrix (t0ij ) as the current travel time matrix. The improved bound IB1
is not updated, since the starting value computed at the root node is already close to the value of
the best known solution, and the computing time of GA, although small, is not negligible.
The B&C and the Dynamic B&C algorithms can also be applied on the basic model (1)-(6),
whose continuous relaxation is strengthened by the SECs, by including all the improved bounds
and the improved value of M . This alternative method will be considered for comparison in the
computational experiments.

6 Computational Results
All the models and algorithms were implemented in C++. All the experiments were executed on
an Intel(R) Core(TM) i7-6900K with 3.20 GHz and 64 GB of RAM (with a single thread), using
GNU/Linux Ubuntu 16.04, and CPLEX 12.7.1 was used as solver of the MIP models and of their
continuous relaxations. In the computational experiments, we considered three sets of instances:

14
 • Set 1: the 22 symmetric instances introduced in Taş et al. (2016), adapted from the TSPLIB
(Reinelt (1991)), containing up to 45 nodes (and an additional node for the ending depot),

• Set 2: all the instances of TSPLIB up to 58 nodes not considered in Set 1 (symmetric) and
up to 45 nodes (asymmetric),

• Set 3: all the instances of TSPLIB up to 200 nodes not considered in Set 2 (symmetric and
asymmetric). Note that asymmetric instances with up to 200 nodes contain at most 170
nodes.

In Taş et al. (2016), the original travel times were modified in order to have the same average
travel time per arc in each instance. The travel times of the new instances were adjusted according
to the same rule used in Taş et al. (2016).
For comparison with Taş et al. (2016) we consider the same linear and quadratic service time
functions, defined as follows:

• small service times: si = 5(10−3 )bi + 3(10−2 ) (i ∈ N \ {0, n + 1});

• medium service times: si = 10−2 bi + 6(10−2 ) (i ∈ N \ {0, n + 1});

• large service times: si = 2(10−2 )bi + 1.2(10−1 ) (i ∈ N \ {0, n + 1});

• quadratic service times: si = 4(10−5 )b2i − 4(10−3 )bi + 3(10−1 ) (i ∈ N \ {0, n + 1}).

We first consider the instances of Set 1, and show in Sections 6.1 and 6.2 the corresponding
results. In Section 6.1, we report the computational results of the GA algorithm on the 22 symmetric
instances introduced in Taş et al. (2016). Then, in Section 6.2, we report the comparison, in terms
of the lower bounds of the continuous relaxation (Section 6.2.1) and of the integer solutions (Section
6.2.2), of the proposed methods with the best method (TGJL16) presented in Taş et al. (2016), by
considering the 22 symmetric instances from Taş et al. (2016). In Section 6.3, we report the results
obtained by GA and by the exact algorithms on the instances of Set 2. Finally, in Section 6.4, we
report the results obtained by GA on the instances of Set 3.
For the B&C and Dynamic B&C algorithms, as well as for TGJL16, we set a time limit of
7200 seconds for the instances of Set 1, and of 50000 seconds for the instances of Set 2. To have a
fair comparison with TGJL16, we implemented the corresponding model (which includes the GG
constraints and the bounds B1 , B2 and B3 ) and run the solver CPLEX 12.7.1 on the same computer.
As mentioned in the introduction, the TSP-TS can be formulated using the layered graph model
proposed in Picard and Queyranne (1978), as a TDTSP. The state-of-the-art method for solving
TDTSP is the Branch-and-Cut-and-Price algorithm proposed in Abeledo et al. (2013): however,
we could not report a comparison with the methods that we propose, since the code is no longer
available from the authors at the time of writing.

6.1 Genetic Algorithm on Instances of Set 1

We report the results obtained by the GA algorithm on the 22 symmetric instances of Set 1
introduced in Taş et al. (2016) with small and medium service times in Table 2, and with large
and quadratic service times in Table 3. The GA algorithm was run 10 times on each instance. In
each table, we show the instance name, in which the number indicates the number of nodes (i.e.,
n + 1) in the corresponding graph, and the best known solution value (obtained by TGJL16 or
by the proposed B&C and Dynamic B&C algorithms). Then, we report for the considered service
time function (small and medium in Table 2, large and quadratic in Table 3), the results obtained

15
by the GA algorithm in two variants: the first one does not consider the separation of the SECs
in the solution of the continuous relaxation, used to generate a subset of the initial population,
while the second one includes it. For each variant and instance, we report the average and the
minimum percentage gaps (computed w.r.t. the best solution value), obtained over the 10 runs,
and the average computing time (expressed in seconds) over the 10 runs. Note that the minimum
percentage gap is obtained by executing 10 runs of the GA algorithm. Thus, the corresponding
computing time is ten times the average computing time (and is not reported in the tables). In the
last two rows, we show the averages of the values in the corresponding columns computed over all
instances and report the number of best known solutions found.
As we can observe, when considering the best out of 10 runs, at least one of the GA algorithms
(without or with SECs) finds the best known solution for all instances. In addition, when consid-
ering the best out of 10 runs, the GA with SECs finds the best known solution for all instances
except two (one with the small and one with the large service time functions). The computing
time is very short in all cases, being at most 1.7 seconds, on average, for the quadratic service
time function with SECs. The average percentage gap over 10 runs when the SECs separation
is included is always not worse than when it is neglected. The minimum percentage gap over 10
runs when the SECs separation is taken into account is not worse than when it is neglected, with
the exception of one instance in the case of the small service time function. We can also observe
that the largest average percentage gap among all instances w.r.t. the best known solution value is
0.97% (for instance dantzig42 and medium service time). In the computation of the values of IB1
and IB3 used in the B&C algorithms, we decided to use the minimum cost solutions obtained by
the GA algorithm with SECs out of 5 runs instead of 10 to limit the computing times.
As shown in Tables 2 and 3, the variant of the GA algorithm that includes SECs is very
effective. To show the importance of using the continuous relaxation (CR) solution to build the
initial population, as explained in Section 5.1, we tested the GA algorithm without CR, i.e., by
building 50% of the initial population by the random route algorithm and 50% by the randomized
NNH algorithm, and compared the obtained results with those of GA with SECs. The detailed
results are reported, in the Appendix (available as Supplementary Material online), for small and
medium service times as well as for large and quadratic service times. The results show that a
smaller number of best solutions and a slightly larger average gap are obtained when CR is not
considered.

6.2 Comparison with TGJL16 on Instances of Set 1

This section is devoted to the comparison of the results obtained by the proposed exact algorithms
with the results obtained by the best method (TGJL16) presented in Taş et al. (2016). We show
in Section 6.2.1 the comparison of the lower bounds obtained by solving the continuous relaxation
of the best model proposed in Taş et al. (2016) with those obtained by solving the continuous
relaxation of the new model NMI, enhanced with the improved bounds and the SECs separation.
In addition, we show in Section 6.2.2 the comparison of the integer solutions obtained by TGJL16
and by the proposed B&C and Dynamic B&C algorithms. The times required for computing M ,
and the lower and upper bounds are included in the total computing times shown in the following
tables, both for TGJL16 and the proposed methods.

6.2.1 Lower Bounds

We present, in Tables 4, 5, 6 and 7, the comparison of the lower bounds obtained by TGJL16
and by the proposed model with or without the improvements of the lower and upper bounds and

16
 Small service times Medium service times
without SECs with SECs without SECs with SECs
# inst Best Avg% Min% Time Avg% Min% Time Best Avg% Min% Time Avg% Min% Time
burma14 228.83 0.00 0.00 0.13 0.00 0.00 0.13 236.44 0.00 0.00 0.10 0.00 0.00 0.11
ulysses16 271.74 0.00 0.00 0.13 0.06 0.00 0.13 279.57 0.00 0.00 0.13 0.02 0.00 0.13
gr17 238.39 0.00 0.00 0.15 0.00 0.00 0.15 245.40 0.00 0.00 0.15 0.00 0.00 0.15
gr21 237.11 0.00 0.00 0.21 0.00 0.00 0.22 249.32 0.00 0.00 0.21 0.00 0.00 0.22
ulysses22 306.43 0.01 0.00 0.23 0.00a 0.00 0.23 318.06 0.16 0.00 0.22 0.38 0.00 0.23
gr24 269.09 0.00 0.00 0.27 0.00 0.00 0.30 284.93 0.00 0.00 0.27 0.00 0.00 0.29
fri26 247.99 0.00 0.00 0.33 0.00 0.00 0.33 263.01 0.00 0.00 0.33 0.00 0.00 0.34
bayg29 345.49 0.30 0.00 0.41 0.05 0.00 0.43 371.22 0.23 0.00 0.41 0.00 0.00 0.43
bays29 309.27 0.26 0.00 0.41 0.16 0.00 0.42 331.90 0.20 0.00 0.41 0.15 0.00 0.42
att30 253.85 0.00 0.00 0.44 0.00 0.00 0.46 273.10 0.00 0.00 0.44 0.00 0.00 0.45
dantzig30 324.21 0.54 0.00 0.43 0.21 0.00 0.45 349.60 0.30 0.30 0.43 0.09 0.00 0.45
eil30 323.40 0.38 0.00 0.45 0.25 0.00 0.46 349.16 0.34 0.00 0.44 0.17 0.00 0.45
gr30 283.91 0.00 0.00 0.46 0.00 0.00 0.48 305.23 0.11 0.00 0.45 0.00 0.00 0.46

17
hk30 324.20 0.50 0.00 0.46 0.00 0.00 0.48 347.35 0.63 0.00 0.45 0.00 0.00 0.47
swiss30 342.50 0.00 0.00 0.46 0.00 0.00 0.47 366.78 0.00 0.00 0.46 0.00 0.00 0.47
eil35 363.38 0.34 0.00 0.65 0.11 0.00 0.68 397.42 0.34 0.00 0.64 0.16 0.00 0.68
gr35 281.82 0.00 0.00 0.68 0.00 0.00 0.74 306.91 0.00 0.00 0.68 0.00 0.00 0.75
swiss35 373.60 0.00 0.00 0.71 0.00 0.00 0.73 406.92 0.00 0.00 0.70 0.00 0.00 0.73
eil40 410.35 0.18 0.00 0.91 0.02 0.00 0.96 452.89 0.33 0.00 0.90 0.16 0.00 0.95
dantzig42 257.37 0.33 0.00 1.02 0.59 0.00 1.08 285.07 0.72 0.00 1.00 0.97 0.00 1.06
swiss42 351.15 0.03 0.00 1.04 0.03 0.00 1.07 388.64 0.59 0.00 0.99 0.07 0.00 1.05
eil45 448.11 0.03 0.00 1.24 0.07 0.04 1.30 502.52 0.16 0.00 1.22 0.25 0.00 1.30
Avg. 308.74 0.13 0.00 0.51 0.07 0.00a 0.53 332.34 0.19 0.01 0.50 0.11 0.00 0.53
#Best 11 22 11 21 10 21 12 22

Table 2: GA on instances of Set 1 (Taş et al. (2016)) with small and medium service times.
a
the gap is ≥ 0.001% and ≤ 0.005%
 Large service times Quadratic service times
without SECs with SECs without SECs with SECs
# inst Best Avg% Min% Time Avg% Min% Time Best Avg% Min% Time Avg% Min% Time
burma14 252.62 0.00 0.00 0.10 0.00 0.00 0.11 224.83 0.00 0.00 0.13 0.00 0.00 0.19
ulysses16 296.28 0.00 0.00 0.13 0.00 0.00 0.13 268.14 0.03 0.00 0.15 0.03 0.00 0.30
gr17 260.34 0.00 0.00 0.15 0.00 0.00 0.15 234.82 0.00 0.00 0.18 0.00 0.00 0.37
gr21 275.96 0.00 0.00 0.21 0.00 0.00 0.22 232.77 0.00 0.00 0.27 0.00 0.00 0.47
ulysses22 343.57 0.00 0.00 0.23 0.00 0.00 0.23 301.58 0.00 0.00 0.28 0.17 0.00 0.71
gr24 320.42 0.03 0.00 0.27 0.00 0.00 0.29 263.04 0.00 0.00 0.35 0.00 0.00 1.01
fri26 297.39 0.00 0.00 0.32 0.00 0.00 0.34 239.08 0.00 0.00 0.41 0.00 0.00 0.90
bayg29 430.35 0.14 0.00 0.40 0.05 0.00 0.43 345.11 0.00 0.00 0.51 0.00 0.00 1.39
bays29 383.78 0.16 0.00 0.41 0.16 0.00 0.43 305.46 0.24 0.00 0.52 0.24 0.00 1.24
att30 316.51 0.00 0.00 0.43 0.00 0.00 0.45 246.98 0.00 0.00 0.56 0.00 0.00 1.65
dantzig30 404.54 0.73 0.69 0.41 0.57 0.00 0.44 321.89 0.55 0.55 0.57 0.22 0.00 1.96
eil30 408.23 0.21 0.00 0.44 0.15 0.00 0.44 320.74 0.40 0.00 0.55 0.29 0.00 1.33

18
gr30 353.89 0.00 0.00 0.45 0.00 0.00 0.47 279.94 0.00 0.00 0.58 0.00 0.00 1.35
hk30 400.88 0.75 0.00 0.43 0.00 0.00 0.44 318.63 0.67 0.00 0.58 0.00 0.00 1.84
swiss30 422.54 0.00 0.00 0.46 0.00 0.00 0.47 340.42 0.00 0.00 0.58 0.00 0.00 1.25
eil35 474.90 0.22 0.00 0.53 0.24 0.20 0.59 365.34 0.32 0.00 0.80 0.09 0.00 2.15
gr35 365.75 0.00 0.00 0.67 0.00 0.00 0.70 276.18 0.00 0.00 0.85 0.00 0.00 2.28
swiss35 485.44 0.00 0.00 0.69 0.00 0.00 0.72 378.36 0.00 0.00 0.88 0.00 0.00 2.15
eil40 556.10 0.07 0.00 0.75 0.07 0.00 0.80 421.10 0.07 0.00 1.12 0.10 0.00 3.03
dantzig42 352.36 0.42 0.00 0.93 0.00 0.00 1.03 247.25 0.09 0.00 1.25 0.67 0.00 4.54
swiss42 480.29 0.12 0.00 0.90 0.05 0.00 1.04 350.45 0.26 0.00 1.27 0.03 0.00 3.50
eil45 638.13 0.09 0.00 1.19 0.07 0.00 1.28 474.44 0.12 0.00 1.47 0.27 0.00 4.09
Avg. 387.29 0.13 0.03 0.48 0.06 0.01 0.51 307.12 0.13 0.03 0.63 0.10 0.00 1.71
#Best 11 21 14 21 12 21 12 22

Table 3: GA on instances of Set 1 (Taş et al. (2016)) with large and quadratic service times.
the SECs separation, by considering small, medium, large or quadratic service time functions. In
particular, in each table, we report the results obtained by the following models:
• TGJL16: continuous relaxation of the basic model (1)-(6), enhanced with the GG constraints
(7)-(8), with the computation of the M value, and of the bounds B1 , B2 (only for the linear
service time functions) and B3 , described in Section 3.1;
• NMB+Bs: continuous relaxation of the model (17)-(26), for the linear service time functions,
and of the model (17)-(21), (24)-(26), (29)-(30), for the quadratic service time function, with
the computation of the M value, and of the bounds B1 , B2 (only for the linear service time
functions) and B3 , described in Section 3.1;
• NMI+IBs: continuous relaxation of the model NMI (corresponding to the model (17)-(20),
(23)-(26), (27)-(28), for the linear service time functions, and to the model (17)-(20), (24)-(26),
(27), (30)-(31), for the quadratic service time function) including in the model the improved
bounds IB1 , IB2 (only for linear service time functions) and IB3 , described in Section 4.1;
• NMI+IBs+SECs: NMI+IBs, combined with the separation of the SECs (described in Section
4.2);
• basic+IBs+SECs: continuous relaxation of model (1)-(6), with the computation of the im-
proved M value described in Section 4.1, including in the model the improved bounds IB1 ,
IB2 (only for linear service time functions) and IB3 , described in Section 4.1, and the sepa-
ration of the SECs (described in Section 4.2).
In each table, we report, for each instance, the instance name and the best known solution value.
Then, for each considered model, we report the lower bound and the corresponding computing
time (expressed in seconds). In the last row, we display the averages of the values reported in the
corresponding columns.
By looking at the results, we can see that the lower bounds obtained by NMB+Bs are worse
than those obtained by TGJL16, since the model NMB is used without any enhancement. When
the improved bounds are inserted, the lower bounds increase significantly and become, on aver-
age, better than those of TGJL16 for all the service time functions. It is evident that using the
exponentially many SECs, which are dynamically separated at the root node, gives another con-
siderable improvement, although it comes at the expenses of larger computing times. However, the
computing times are still short, with the exception of the quadratic service time function, which
makes the TSP-TS problem harder to solve. We can also observe that if we use the basic model
instead of NMI, and include all the proposed improvements, the obtained lower bounds are slightly
worse, although the computing times are larger for NMI than for the basic model. However, as
it will be shown in Section 6.2.2, both the B&C and the Dynamic B&C algorithms have a better
performance when model NMI is used. Finally, we can also see that the lower bounds obtained
by NMI+IBs+SECs are rather close to the best known solution values, even though the case of
quadratic service time function shows a larger computing time. One reason for this is that CPLEX
has to solve the continuous relaxation of a quadratic model instead of a linear one. Overall, we can
conclude that both the improved bounds and the separation of the exponentially many SECs help
to find better lower bounds in all cases. In particular, by comparing the percentage gap between
the average of the best known solution values and the average of the lower bound values obtained
by TGJL16 and by NMI+IBs+SECs, which obtains the best lower bound values, we can note that:
the percentage gap is reduced from 13.69% to 2.25% for the small service times, from 15.10% to
4.23% for the medium ones, from 17.65% to 7.91% for the large ones, and from 16.53% to 6.20%
for the quadratic ones.

19
 TGJL16 NMB+Bs NMI+IBs NMI+IBs+SECs basic+IBs+SECs
# inst Best LB Time LB Time LB Time LB Time LB Time
burma14 228.83 204.04 0.07 190.25 0.01 206.02 0.05 226.28 0.05 225.65 0.04
ulysses16 271.74 235.39 0.00 221.70 0.01 237.97 0.05 268.66 0.05 267.83 0.04
gr17 238.39 198.86 0.00 189.73 0.01 207.35 0.04 237.29 0.05 236.27 0.03
gr21 237.11 213.89 0.01 208.98 0.01 216.81 0.07 233.48 0.06 232.75 0.04
ulysses22 306.43 247.10 0.00 227.90 0.02 251.49 0.05 299.86 0.07 297.53 0.04
gr24 269.09 234.59 0.01 222.27 0.02 235.55 0.08 265.12 0.20 265.12 0.04
fri26 247.99 220.17 0.01 216.87 0.02 221.74 0.12 244.59 0.18 244.59 0.05
bayg29 345.49 315.78 0.02 307.47 0.02 314.93 0.14 339.71 0.23 339.71 0.05
bays29 309.27 277.44 0.02 268.42 0.02 276.65 0.15 302.66 0.21 302.66 0.06
att30 253.85 194.16 0.01 178.06 0.02 191.55 0.12 246.30 0.16 245.95 0.06
dantzig30 324.21 271.53 0.01 232.61 0.02 271.68 0.16 316.13 0.13 313.39 0.06
eil30 323.40 295.20 0.02 291.92 0.02 295.30 0.16 317.34 0.31 317.34 0.06
gr30 283.91 238.55 0.02 231.10 0.02 240.07 0.20 276.92 0.19 276.44 0.06
hk30 324.20 272.37 0.01 258.26 0.02 269.43 0.14 317.09 0.19 316.70 0.06
swiss30 342.50 308.29 0.02 302.84 0.02 309.66 0.16 335.10 0.18 334.52 0.06
eil35 363.38 329.39 0.04 325.34 0.03 329.16 0.23 352.72 0.70 352.72 0.09
gr35 281.82 229.32 0.03 219.34 0.03 229.53 0.20 273.48 0.31 273.07 0.09
swiss35 373.60 318.26 0.04 314.53 0.03 319.18 0.23 363.35 0.35 363.17 0.08
eil40 410.35 365.76 0.05 359.75 0.06 365.11 0.30 398.99 0.97 398.99 0.11
dantzig42 257.37 213.92 0.03 193.50 0.05 214.08 0.41 248.58 0.50 248.06 0.12
swiss42 351.15 280.76 0.03 273.90 0.05 282.84 0.33 340.61 0.92 340.61 0.13
eil45 448.11 397.54 0.06 389.35 0.09 396.72 0.45 435.40 1.81 435.40 0.15
Avg. 308.74 266.47 0.02 255.64 0.03 267.40 0.17 301.80 0.36 301.29 0.07

Table 4: Comparison of Lower Bounds with small service times on instances of Set 1 (Taş et al.
(2016)).

TGJL16 NMB+Bs NMI+IBs NMI+IBs+SECs basic+IBs+SECs

# inst Best LB Time LB Time LB Time LB Time LB Time
burma14 236.44 207.89 0.01 193.98 0.01 210.33 0.05 231.08 0.05 229.91 0.04
ulysses16 279.57 239.02 0.01 225.21 0.01 242.11 0.04 272.98 0.05 272.00 0.03
gr17 245.40 202.78 0.01 193.55 0.01 212.76 0.04 242.89 0.05 241.07 0.03
gr21 249.32 220.70 0.01 215.73 0.01 223.48 0.06 240.92 0.06 240.33 0.04
ulysses22 318.06 252.18 0.01 232.83 0.02 257.26 0.05 305.53 0.07 303.14 0.04
gr24 284.93 245.19 0.02 232.77 0.02 246.72 0.09 276.64 0.13 276.64 0.04
fri26 263.01 229.06 0.02 225.74 0.02 232.89 0.10 255.83 0.17 255.83 0.05
bayg29 371.22 335.54 0.02 327.16 0.03 334.48 0.13 359.48 0.21 359.48 0.05
bays29 331.90 293.82 0.02 284.71 0.02 292.77 0.15 319.04 0.18 319.04 0.06
att30 273.10 204.06 0.01 187.90 0.02 201.68 0.15 256.90 0.17 256.66 0.06
dantzig30 349.60 285.16 0.01 246.05 0.01 284.92 0.16 330.72 0.14 327.90 0.06
eil30 349.16 316.39 0.02 313.07 0.02 316.31 0.13 338.53 0.26 338.53 0.06
gr30 305.23 250.89 0.02 243.37 0.02 252.78 0.13 289.96 0.20 289.56 0.06
hk30 347.35 286.35 0.02 272.15 0.02 283.90 0.15 331.72 0.17 331.46 0.06
swiss30 366.78 321.90 0.01 316.40 0.01 324.80 0.13 350.37 0.18 349.89 0.06
eil35 397.42 356.91 0.03 352.82 0.04 356.56 0.21 380.26 0.71 380.26 0.09
gr35 306.91 244.39 0.03 234.33 0.03 244.20 0.20 288.72 0.41 288.39 0.09
swiss35 406.92 337.99 0.03 334.23 0.03 340.45 0.24 384.76 0.38 384.67 0.08
eil40 452.89 400.83 0.05 394.77 0.06 400.00 0.38 434.06 0.81 434.06 0.10
dantzig42 285.07 230.60 0.03 210.07 0.05 230.67 0.45 266.17 0.51 265.77 0.12
swiss42 388.64 304.39 0.03 297.46 0.04 307.74 0.32 366.10 0.96 366.10 0.12
eil45 502.52 441.73 0.06 433.45 0.10 440.49 0.44 479.59 1.59 479.59 0.15
Avg. 332.34 282.17 0.02 271.26 0.03 283.51 0.17 318.28 0.34 317.74 0.07

Table 5: Comparison of Lower Bounds with medium service times on instances of Set 1 (Taş et al.
(2016)).

20
 TGJL16 NMB+Bs NMI+IBs NMI+IBs+SECs basic+IBs+SECs
# inst Best LB Time LB Time LB Time LB Time LB Time
burma14 252.62 216.01 0.01 201.89 0.00 218.56 0.05 240.11 0.05 238.86 0.04
ulysses16 296.28 246.74 0.01 232.71 0.00 250.69 0.04 281.55 0.05 280.82 0.03
gr17 260.34 211.21 0.01 201.81 0.00 223.01 0.04 253.88 0.05 251.31 0.03
gr21 275.96 235.57 0.01 230.53 0.01 239.67 0.06 257.23 0.06 256.78 0.04
ulysses22 343.57 263.27 0.01 243.66 0.01 268.09 0.05 317.25 0.08 315.30 0.04
gr24 320.42 268.84 0.01 256.22 0.01 271.69 0.09 302.38 0.13 302.38 0.05
fri26 297.39 248.99 0.01 245.62 0.01 258.26 0.11 281.36 0.17 281.36 0.05
bayg29 430.35 380.70 0.02 372.16 0.02 379.12 0.14 404.63 0.22 404.63 0.05
bays29 383.78 331.23 0.02 321.95 0.02 329.66 0.14 356.44 0.23 356.44 0.06
att30 316.51 226.76 0.01 210.46 0.01 225.01 0.14 281.20 0.18 281.09 0.06
dantzig30 404.54 316.18 0.01 276.68 0.02 313.91 0.13 362.62 0.14 360.77 0.06
eil30 408.23 365.14 0.02 361.74 0.02 364.70 0.14 387.28 0.30 387.28 0.06
gr30 353.89 279.03 0.02 271.39 0.02 281.24 0.12 319.57 0.23 319.29 0.05
hk30 400.88 318.33 0.02 303.98 0.02 317.08 0.14 365.34 0.16 365.22 0.06
swiss30 422.54 352.92 0.01 347.34 0.01 359.11 0.12 384.96 0.17 384.66 0.06
eil35 474.90 421.85 0.04 417.68 0.03 421.23 0.25 445.25 0.54 445.25 0.09
gr35 365.75 279.71 0.03 269.49 0.02 278.94 0.21 324.30 0.48 324.09 0.08
swiss35 485.44 384.08 0.03 380.27 0.02 389.97 0.30 434.59 0.39 434.59 0.09
eil40 556.10 485.75 0.07 479.58 0.06 484.57 0.37 518.99 0.97 518.99 0.12
dantzig42 352.36 271.00 0.03 250.27 0.04 271.20 0.36 308.58 0.46 308.41 0.11
swiss42 480.29 361.62 0.04 354.55 0.04 368.08 0.38 427.52 0.96 427.52 0.12
eil45 638.13 551.55 0.07 543.11 0.07 549.56 0.44 589.41 1.74 589.41 0.15
Avg. 387.29 318.93 0.02 307.87 0.02 321.06 0.17 356.57 0.35 356.11 0.07

Table 6: Comparison of Lower Bounds with large service times on instances of Set 1 (Taş et al.
(2016)).

TGJL16 NMB+Bs NMI+IBs NMI+IBs+SECs basic+IBs+SECs

# inst Best LB Time LB Time LB Time LB Time LB Time
burma14 224.83 200.33 0.04 186.20 0.05 201.89 0.12 221.93 0.55 221.54 0.49
ulysses16 268.14 231.91 0.05 217.68 0.05 233.41 0.13 264.50 1.39 263.87 0.81
gr17 234.82 195.12 0.04 185.64 0.05 201.74 0.13 232.14 1.94 231.68 0.99
gr21 232.77 207.48 0.06 202.47 0.13 210.05 0.35 226.50 2.36 225.58 1.76
ulysses22 301.58 242.31 0.07 221.60 0.14 245.49 0.33 293.87 5.80 292.29 2.54
gr24 263.04 224.71 0.08 211.55 0.20 224.12 0.42 255.55 6.24 254.40 4.56
fri26 239.08 211.92 0.10 208.46 0.25 211.75 0.50 235.09 8.58 234.25 4.38
bayg29 345.11 297.66 0.12 288.43 0.32 297.52 0.59 322.72 15.62 321.60 7.46
bays29 305.46 262.43 0.12 252.48 0.35 261.74 0.71 288.22 11.25 287.64 6.61
att30 246.98 185.09 0.15 168.01 0.41 181.79 0.56 236.94 16.63 236.12 10.68
dantzig30 321.89 259.00 0.15 216.58 0.41 257.48 0.60 303.84 17.85 300.34 12.84
eil30 320.74 275.86 0.12 272.25 0.39 276.50 0.59 299.43 10.25 298.00 7.64
gr30 279.94 227.24 0.14 219.08 0.37 229.37 0.56 267.25 16.17 264.36 7.19
hk30 318.63 259.56 0.13 244.35 0.37 256.74 0.53 306.28 17.21 303.18 11.84
swiss30 340.42 295.79 0.14 289.72 0.39 296.40 0.64 323.65 11.17 320.35 6.97
eil35 365.34 304.68 0.22 300.03 0.68 304.80 0.88 328.98 30.96 328.00 17.04
gr35 276.18 215.71 0.24 204.69 0.61 216.45 1.20 261.99 34.68 259.20 15.55
swiss35 378.36 300.50 0.27 296.40 0.72 300.73 1.04 346.96 23.35 343.68 16.10
eil40 421.10 335.17 0.34 328.11 0.96 334.69 1.44 369.47 50.31 368.00 28.89
dantzig42 247.25 199.25 0.42 177.45 1.08 197.93 1.19 233.94 61.28 232.35 47.28
swiss42 350.45 259.97 0.39 252.34 1.28 260.23 1.09 319.94 59.89 318.00 34.79
eil45 474.44 360.15 0.46 350.78 1.42 358.26 1.31 398.72 27.68 397.00 44.77
Avg. 307.12 252.36 0.18 240.65 0.48 252.69 0.68 288.09 19.60 286.43 13.24

Table 7: Comparison of Lower Bounds with quadratic service times on instances of Set 1 (Taş et al.
(2016)).

21
6.2.2 Integer Solutions
Tables 8, 9, 10 and 11 present the comparison of the solutions obtained by TGJL16 and by the B&C
and Dynamic B&C algorithms, with small, medium, large or quadratic service time functions. The
results obtained by the Dynamic B&C algorithm are shown in the cases of applying the algorithm
to both the NMI and the basic models. We consider the following algorithms:

• TGJL16: the basic model (1)-(6), enhanced with the GG constraints (7)-(8), with the com-
putation of the M value and the bounds B1 , B2 (only for the linear service time functions)
and B3 , solved by CPLEX;

• NMI B&C: the model (17)-(20), (23)-(26), (27)-(28), for the linear service time functions, and
the model (17)-(20), (24)-(26), (27), (30)-(31), for the quadratic service time function, with
the computation of the improved M value and of the improved bounds IB1 , IB2 (only for
the linear service time functions) and IB3 (described in Section 4.1), and with the separation
of the SECs (described in Section 4.2), solved by using the CPLEX callback functions for
separating the SECs at every node of the decision tree;

• NMI Dyn-B&C: the method NMI B&C in which, at every node of the decision tree, IB2
(only for the linear service time functions) and the shortest path computations for IB3 are
dynamically updated by considering the x variables fixed by the branching;

• basic Dyn-B&C: as in the NMI Dyn-B&C algorithm, but using the basic model (1)-(6).

For all the instances of Set 1, we consider for each method a time limit of 7200 seconds. We
remark that the available feasible solution obtained by GA is not used as a MIP start, but the
upper bound IB1 +  (with  a very small positive number) is used as a cut-off value for NMI
B&C, NMI Dyn-B&C and basic Dyn-B&C. Similarly, the B1 +  is used as a cut-off value for
TGJL16. We also executed our best solution method NMI Dyn-B&C on all the instances of Set 1
(with small, medium, large and quadratic service times) by using the GA solution as MIP start.
Similar performances were obtained in both cases (MIP start or cut-off value): in particular, when
the GA solution is used as MIP start, the average optimality gap was slightly larger, and, for the
case of medium service times, one less instance was solved to optimality within the time limit.
Therefore, we do not report the computational results obtained by NMI Dyn-B&C with MIP start,
and consider, in all the experiments, the GA solution value as a cut-off value. In addition, we
underline that, for the quadratic service time case, the model becomes quadratic, but the proposed
algorithms are exact also in this case: indeed, CPLEX is used as a quadratic programming solver to
solve the continuous relaxation of the obtained quadratic programming model, the SECs procedure
is applied on the continuous relaxed model, and the branching is applied directly by CPLEX on
the quadratic programming model.
In each table, we report, for each instance, the instance name and the best known solution
value. Then we show, for each method, the best integer solution value found during the execution,
the optimality percentage gap (i.e., the percentage gap between the best upper bound and the
best lower bound found at the end of the solving process), and the corresponding computing time
(expressed in seconds). At the bottom of each table, we report the averages of the values shown
in the corresponding columns, except for the computing time: indeed, not all methods solve to
optimality within the time limit the same subset of instances. Therefore, we report, in row AvgS,
the average computing time over the subset of instances solved to optimality within the time limit
by the corresponding method and, to have a fair comparison, we also show the average computing
time by considering only the instances solved to optimality by a subset of methods. More precisely,

22
we call: (i) Avg. TGJL16 the average computing time over the subset of instances solved by
TGJL16; (ii) Avg. NMI B&C the average computing time over the subset of instances solved
by the NMI B&C algorithm; (iii) Avg. basic Dyn-B&C the average computing time over the
subset of instances solved by the basic Dyn-B&C algorithm. Note that not all these averages are
reported in every table, since, in some cases, the subsets of instances solved by the various methods
coincide. In row AvgS, we also report, in column UB, the average number of nodes explored by
each method computed on the instances solved to optimality within the time limit. The numbers
of nodes explored by each method for each instance can be found in the Appendix (available as
Supplementary Material online) for the small and medium service times, and for the large and
quadratic ones. In addition, we show at the bottom of each table, the number of instances solved,
by each method, to optimality within the time limit. For the case of the quadratic service time
function (Table 11), we show an additional row, called Avg. Feas., since TGJL16 is not able to
find a feasible solution for a subset of instances.
In Table 8, in which the small service times are considered, we can see that all methods solve to
optimality within the time limit all the 22 instances, and the computing times of the proposed B&C
and Dynamic B&C algorithms are about one order of magnitude shorter than those of TGJL16.
In addition, the average numbers of nodes explored by NMI B&C and NMI Dyn-B&C are more
than two orders of magnitude smaller than those explored by TGJL16, which is congruent with the
shorter computing times and the better lower bounds achieved by the proposed methods.
When the medium service times are considered (Table 9), not all the instances can be solved
to optimality within the time limit, and the Dynamic B&C algorithm based on NMI obtains
the largest number of proved optimal solutions (20 out of 22), while TGJL16 can solve only 14
instances to optimality within the time limit. In addition, the average percentage gaps of the
proposed algorithms are always very small. By looking at the average computing times, we can
see that, on the subset of 14 instances solved to optimality within the time limit by all methods,
the fastest one is the Dynamic B&C algorithm based on the basic model. On the contrary, if we
consider the subset of instances solved to optimality within the time limit by NMI B&C, the fastest
algorithm turns out to be the Dynamic B&C based on the NMI formulation. The same happens
if we consider the subset of instances solved by the basic Dynamic B&C algorithm. Also in this
case, the average numbers of nodes explored by NMI B&C and NMI Dyn-B&C are more than two
orders of magnitude smaller than those explored by TGJL16.
In the case of the large service times (Table 10), the TSP-TS instances become harder for all
methods: TGJL16 can solve to optimality within the time limit only 6 instances, and the largest
number of instances solved to optimality within the time limit is 11 (obtained by the Dynamic B&C
algorithm based on the model NMI). As for the medium service times, also in this case, the fastest
algorithm on the subset of instances solved to optimality within the time limit by all methods is
the Dynamic B&C algorithm based on the basic model, while, when we consider the subsets of
instances solved to optimality within the time limit by the NMI B&C and by the basic Dyn-B&C
algorithms, the Dynamic B&C algorithm based on the NMI formulation is the fastest one. In the
latter cases, we can observe that the computing times are significantly reduced by using the NMI
Dyn-B&C algorithm. Note that, since the 9 instances solved to optimality within the time limit
by the NMI B&C algorithm and the 9 instances solved to optimality within the time limit by the
basic Dyn-B&C algorithm are not the same, we do not have the corresponding average computing
time values. In addition, we can observe that the average percentage gap obtained by the NMI
Dyn-B&C algorithm is 2.59%, and is the smallest one. The average numbers of nodes explored by
NMI B&C and NMI Dyn-B&C are about three times smaller than those explored by TGJL16.
Finally, in Table 11, we report the results obtained for the quadratic service time function. Both
the B&C and the Dynamic B&C algorithms based on the NMI formulation are able to determine

23
the same subset of 18 instances solved to optimality within the time limit in similar computing
times. TGJL16 solves 9 instances to optimality within the time limit in an average computing
time that is significantly larger than that of the other algorithms. If we consider the subset of
instances solved to optimality within the time limit by the Dynamic B&C algorithm based on the
basic model and its computing time, we can see that the computing times of both the NMI B&C
and the NMI Dyn-B&C algorithms are considerably shorter. Moreover, the average numbers of
nodes explored by NMI B&C and NMI Dyn-B&C are one order of magnitude smaller than those
explored by TGJL16.
We can conclude that the proposed algorithms outperform TGJL16 in terms of number of
instances solved to optimality within the time limit and computing times. Among the proposed
algorithms, the Dynamic B&C algorithm based on the NMI formulation has globally the best
performance. In particular, with respect to the NMI B&C algorithm, that does not include the
dynamic update of the improved bounds, the NMI Dyn-B&C algorithm is always able to solve a
larger (or equal) number of instances to optimality within the time limit (in shorter or comparable
computing times). This highlights that the dynamic update of the bounds is effective. By comparing
the Dynamic B&C algorithm based on the model NMI with that based on the basic model, we can
see that again the former solves a larger number of instances to optimality within the time limit, and
the computing times of the former are significantly shorter than those of the latter, thus showing
the usefulness of the NMI formulation.

6.3 Genetic and Exact Algorithms on Instances of Set 2

In this section, we consider the instances of Set 2. In Section 6.3.1 we report the results obtained on
13 larger size symmetric instances, corresponding to all the symmetric instances with up to 58 nodes
contained in the TSPLIB (except the instances already considered in Taş et al. (2016)). In Section
6.3.2 we show the results obtained on all the 27 asymmetric instances with up to 45 nodes contained
in the TSPLIB. In both cases, we consider small service times (Tables 12 and 15), as well as medium
service times (Tables 13 and 16), and report the results of the following methods: GA+SECs, i.e.,
the GA in which the continuous relaxation of the basic model, combined with the improved bound
IB2 and the SECs separation, is used for building a subset of the initial population; TGJL16 and
NMI Dyn-B&C (defined as in Section 6.2.2). For these additional instances, we consider for each
exact method a time limit of 50000 seconds. For the GA+SECs algorithm, we consider 10 runs for
each instance.
In each table, we report the instance name and the best solution value found by the three
considered algorithms. Then, for GA+SECs, we show the minimum value of the solutions obtained
over 10 runs, the average, over the 10 runs, percentage gap w.r.t. the best solution value, the
minimum percentage gap found over the 10 runs and the average computing time over the 10 runs.
In addition, we show, for the TGJL16 and the NMI Dyn-B&C algorithms, the lower bound value
of the continuous relaxation at the root node and the corresponding computing time, the integer
solution value obtained at the end of the solving process, the optimality percentage gap (i.e., the
percentage gap between the best upper bound and the best lower bound found at the end of the
solving process), and the corresponding computing time. All the computing times are expressed
in seconds. At the bottom of each table, we display the averages of the values reported in the
corresponding columns, except for the average computing time that is shown, in row AvgS, on the
instances solved to optimality within the time limit by the corresponding method, and separately
for comparison on the instances solved to optimality within the time limit by both exact methods.
In row AvgS, we also report the average number of nodes explored by each method computed
on the instances solved to optimality within the time limit. The numbers of nodes explored by

24
 TGJL16 NMI B&C NMI Dyn-B&C basic Dyn-B&C
# inst Best UB Gap% Time UB Gap% Time UB Gap% Time UB Gap% Time
burma14 228.83 228.83 0.00 0.32 228.83 0.00 0.31 228.83 0.00 0.37 228.83 0.00 0.19
ulysses16 271.74 271.74 0.00 2.18 271.74 0.00 0.41 271.74 0.00 0.33 271.74 0.00 0.30
gr17 238.39 238.39 0.00 1.44 238.39 0.00 0.25 238.39 0.00 0.33 238.39 0.00 0.19
gr21 237.11 237.11 0.00 0.38 237.11 0.00 0.43 237.11 0.00 0.33 237.11 0.00 0.37
ulysses22 306.43 306.43 0.00 19.84 306.43 0.00 1.29 306.43 0.00 1.29 306.43 0.00 1.93
gr24 269.09 269.09 0.00 1.99 269.09 0.00 0.91 269.09 0.00 0.96 269.09 0.00 0.67
fri26 247.99 247.99 0.00 2.62 247.99 0.00 0.83 247.99 0.00 0.83 247.99 0.00 0.56
bayg29 345.49 345.49 0.00 8.33 345.49 0.00 2.15 345.49 0.00 2.42 345.49 0.00 1.70
bays29 309.27 309.27 0.00 18.85 309.27 0.00 3.64 309.27 0.00 4.49 309.27 0.00 4.51
att30 253.85 253.85 0.00 223.61 253.85 0.00 2.50 253.85 0.00 3.37 253.85 0.00 9.95
dantzig30 324.21 324.21 0.00 165.18 324.21 0.00 2.04 324.21 0.00 2.34 324.21 0.00 6.11
eil30 323.40 323.40 0.00 3.42 323.40 0.00 1.94 323.40 0.00 2.02 323.40 0.00 1.22
gr30 283.91 283.91 0.00 14.00 283.91 0.00 1.30 283.91 0.00 1.37 283.91 0.00 1.89

25
hk30 324.20 324.20 0.00 168.18 324.20 0.00 2.30 324.20 0.00 2.46 324.20 0.00 11.25
swiss30 342.50 342.50 0.00 7.65 342.50 0.00 1.36 342.50 0.00 1.41 342.50 0.00 1.20
eil35 363.38 363.38 0.00 44.19 363.38 0.00 29.57 363.38 0.00 32.39 363.38 0.00 11.31
gr35 281.82 281.82 0.00 149.94 281.82 0.00 3.30 281.82 0.00 3.54 281.82 0.00 21.83
swiss35 373.60 373.60 0.00 26.10 373.60 0.00 2.35 373.60 0.00 2.67 373.60 0.00 7.97
eil40 410.35 410.35 0.00 217.95 410.35 0.00 69.54 410.35 0.00 82.15 410.35 0.00 49.87
dantzig42 257.37 257.37 0.00 2302.25 257.37 0.00 35.06 257.37 0.00 36.49 257.37 0.00 130.97
swiss42 351.15 351.15 0.00 2073.20 351.15 0.00 14.00 351.15 0.00 16.10 351.15 0.00 140.62
eil45 448.10 448.10 0.00 2062.93 448.10 0.00 434.35 448.10 0.00 587.63 448.10 0.00 534.56
Avg. 308.74 308.74 0.00 308.74 0.00 308.74 0.00 308.74 0.00
AvgS 104641 341.57 4139 27.72 2705 35.70 8964 42.69
#Opt. 22 22 22 22

Table 8: Comparison of Integer Solutions with small service times on instances of Set 1 (Taş et al. (2016)).
 TGJL16 NMI B&C NMI Dyn-B&C basic Dyn-B&C
# inst Best UB Gap% Time UB Gap% Time UB Gap% Time UB Gap% Time
burma14 236.44 236.44 0.00 1.02 236.44 0.00 0.35 236.44 0.00 0.35 236.44 0.00 0.19
ulysses16 279.57 279.57 0.00 5.16 279.57 0.00 0.42 279.57 0.00 0.50 279.57 0.00 0.53
gr17 245.40 245.40 0.00 1.59 245.40 0.00 0.24 245.40 0.00 0.24 245.40 0.00 0.24
gr21 249.32 249.32 0.00 1.81 249.32 0.00 0.49 249.32 0.00 0.60 249.32 0.00 0.31
ulysses22 318.06 318.06 0.00 268.13 318.06 0.00 3.75 318.06 0.00 4.58 318.06 0.00 35.27
gr24 284.93 284.93 0.00 33.75 284.93 0.00 1.38 284.93 0.00 1.48 284.93 0.00 2.18
fri26 263.01 263.01 0.00 110.43 263.01 0.00 1.32 263.01 0.00 1.37 263.01 0.00 1.48
bayg29 371.22 371.22 0.00 126.62 371.22 0.00 21.95 371.22 0.00 43.84 371.22 0.00 31.50
bays29 331.90 331.90 0.00 551.96 331.90 0.00 90.11 331.90 0.00 82.24 331.90 0.00 119.10
att30 273.10 273.10 3.29 7200.00 273.10 0.00 136.82 273.10 0.00 130.81 273.10 0.00 3201.08
dantzig30 349.60 349.60 2.75 7200.00 349.60 0.00 44.95 349.60 0.00 60.86 349.60 0.00 385.31
eil30 349.16 349.16 0.00 22.01 349.16 0.00 11.38 349.16 0.00 10.86 349.16 0.00 6.10
gr30 305.23 305.23 0.00 324.07 305.23 0.00 5.11 305.23 0.00 5.52 305.23 0.00 34.00
hk30 347.35 347.35 1.05 7200.00 347.35 0.00 21.64 347.35 0.00 24.34 347.35 0.00 823.23

26
swiss30 366.78 366.78 0.00 350.44 366.78 0.00 6.37 366.78 0.00 6.95 366.78 0.00 9.95
eil35 397.42 397.42 0.00 2139.22 397.42 0.00 4241.00 397.42 0.00 1310.73 397.42 0.00 280.15
gr35 306.91 306.91 2.77 7200.00 306.91 0.00 81.35 306.91 0.00 81.46 306.91 0.00 2414.11
swiss35 406.92 406.92 0.00 3680.01 406.92 0.00 67.37 406.92 0.00 52.19 406.92 0.00 348.97
eil40 452.89 452.89 1.24 7200.00 452.89 1.19 7200.00 452.89 0.00 6087.80 452.89 0.00 1913.40
dantzig42 285.07 285.07 5.81 7200.00 285.07 1.77 7200.00 285.07 0.80 7200.00 285.07 3.37 7200.00
swiss42 388.64 388.64 5.26 7200.00 388.64 1.88 7200.00 388.64 0.00 6825.77 388.64 3.48 7200.00
eil45 502.52 502.52 3.09 7200.00 502.52 2.77 7200.00 502.52 2.13 7200.00 502.52 2.40 7200.00
Avg. 332.34 332.34 1.15 332.34 0.35 332.34 0.13 332.34 0.42
AvgS 220522 544.02 35388 263.11 37690 736.63 139823 505.64
#Opt. 14 18 20 19
Avg. TGJL16 544.02 317.95 108.68 62.14
Avg. NM B&C 263.11 101.05 427.43
Avg. basic Dyn-B&C 416.14 505.64

Table 9: Comparison of Integer Solutions with medium service times on instances of Set 1 (Taş et al. (2016)).
 TGJL16 NMI B&C NMI Dyn-B&C basic Dyn-B&C
# inst Best UB Gap% Time UB Gap% Time UB Gap% Time UB Gap% Time
burma14 252.62 252.62 0.00 2.66 252.62 0.00 0.45 252.62 0.00 0.48 252.62 0.00 0.52
ulysses16 296.28 296.28 0.00 45.86 296.28 0.00 1.11 296.28 0.00 1.36 296.28 0.00 3.40
gr17 260.34 260.34 0.00 7.26 260.34 0.00 0.49 260.34 0.00 0.49 260.34 0.00 0.52
gr21 275.96 275.96 0.00 13.82 275.96 0.00 1.88 275.96 0.00 1.83 275.96 0.00 1.08
ulysses22 343.57 343.57 4.83 7200.00 343.57 0.00 58.29 343.57 0.00 46.59 343.57 0.00 1776.91
gr24 320.42 320.42 0.00 1285.79 320.42 0.00 35.36 320.42 0.00 26.27 320.42 0.00 70.38
fri26 297.39 297.39 4.03 7200.00 297.39 0.00 20.74 297.39 0.00 21.93 297.39 0.00 62.84
bayg29 430.35 430.35 2.82 7200.00 430.35 1.94 7200.00 430.35 0.00 6926.77 430.35 0.80 7200.00
bays29 383.78 383.78 4.72 7200.00 383.78 4.04 7200.00 383.78 3.01 7200.00 383.78 4.03 7200.00
att30 316.51 316.51 10.42 7200.00 316.51 6.78 7200.00 316.51 5.35 7200.00 316.51 7.68 7200.00
dantzig30 404.54 404.54 9.67 7200.00 404.54 4.24 7200.00 404.54 3.29 7200.00 404.54 7.06 7200.00
eil30 408.23 408.23 0.00 2437.02 408.23 0.00 5455.40 408.23 0.00 798.22 408.23 0.00 301.15
gr30 353.89 353.89 6.82 7200.00 353.89 0.00 5785.61 353.89 0.00 1978.32 353.89 2.19 7200.00
hk30 400.88 400.88 8.29 7200.00 400.88 3.20 7200.00 400.88 2.62 7200.00 400.88 5.99 7200.00

27
swiss30 422.54 422.54 6.77 7200.00 422.54 1.74 7200.00 422.54 0.00 1616.20 422.54 0.00 6408.20
eil35 474.90 474.90 3.87 7200.00 474.90 4.14 7200.00 474.90 3.19 7200.00 474.90 3.19 7200.00
gr35 365.75 365.75 10.45 7200.00 365.75 7.06 7200.00 365.75 5.37 7200.00 365.75 9.02 7200.00
swiss35 485.44 485.44 9.84 7200.00 485.44 6.95 7200.00 485.44 5.33 7200.00 485.44 7.31 7200.00
eil40 556.10 556.10 5.19 7200.00 556.10 5.32 7200.00 556.10 4.68 7200.00 556.10 4.62 7200.00
dantzig42 352.36 352.36 12.55 7200.00 352.36 10.61 7200.00 352.36 9.43 7200.00 352.36 10.61 7200.00
swiss42 480.29 480.29 11.77 7200.00 480.29 9.18 7200.00 480.29 8.31 7200.00 480.29 9.14 7200.00
eil45 638.13 638.13 6.81 7200.00 638.13 6.68 7200.00 638.13 6.41 7200.00 638.13 6.00 7200.00
Avg. 387.29 387.29 5.40 387.29 3.27 387.29 2.59 387.29 3.53
AvgS 362003 632.07 125917 1262.15 117217 1038.04 251471 958.33
#Opt. 6 9 11 9
Avg. TGJL16 632.07 915.78 138.11 62.84
Avg. NM B&C 1262.15 319.50
Avg. basic Dyn-B&C 279.26 958.33

Table 10: Comparison of Integer Solutions with large service times on instances of Set 1 (Taş et al. (2016)).
 TGJL16 NMI B&C NMI Dyn-B&C basic Dyn-B&C
# inst Best UB Gap% Time UB Gap% Time UB Gap% Time UB Gap% Time
burma14 224.83 224.83 0.00 2.79 224.83 0.00 3.64 224.83 0.00 3.65 224.83 0.00 1.09
ulysses16 268.14 268.14 0.00 9.94 268.14 0.00 4.25 268.14 0.00 4.11 268.14 0.00 1.95
gr17 234.82 234.82 0.00 10.26 234.82 0.00 5.83 234.82 0.00 5.88 234.82 0.00 1.89
gr21 232.77 232.77 0.00 6.58 232.77 0.00 8.18 232.77 0.00 8.00 232.77 0.00 4.29
ulysses22 301.58 301.58 0.00 666.92 301.58 0.00 18.89 301.58 0.00 19.71 301.58 0.00 14.41
gr24 263.04 263.04 0.00 45.21 263.04 0.00 17.90 263.04 0.00 18.04 263.04 0.00 19.75
fri26 239.08 239.08 0.00 28.36 239.08 0.00 21.93 239.08 0.00 21.92 239.08 0.00 25.24
bayg29 345.11 345.11 4.06 7200.00 345.11 0.00 292.44 345.11 0.00 441.98 345.11 1.06 7200.00
bays29 305.46 305.46 2.12 7200.00 305.46 0.00 145.85 305.46 0.00 179.04 305.46 0.00 6409.08
att30 246.98 246.98 1.15 7200.00 246.98 0.00 40.64 246.98 0.00 43.63 246.98 0.00 428.38
dantzig30 321.89 321.89 3.94 7200.00 321.89 0.00 153.15 321.89 0.00 240.26 321.89 0.00 2594.89
eil30 320.74 320.74 2.39 7200.00 320.74 0.00 398.16 320.74 0.00 499.42 320.74 0.00 5381.88
gr30 279.94 279.94 0.00 2913.02 279.94 0.00 42.25 279.94 0.00 44.07 279.94 0.00 90.91
hk30 318.63 318.63 3.16 7200.00 318.63 0.00 49.71 318.63 0.00 52.49 318.63 0.00 2626.19
swiss30 340.42 340.42 0.00 1819.48 340.42 0.00 40.83 340.42 0.00 41.58 340.42 0.00 226.65

28
eil35 365.34 365.34 8.14 7200.00 365.34 4.85 7200.00 365.34 4.61 7200.00 365.34 7.64 7200.00
gr35 276.18 276.18 8.93 7200.00 276.18 0.00 113.79 276.18 0.00 129.96 276.18 3.14 7200.00
swiss35 378.36 387.31 12.86 7200.00 378.36 0.00 1839.16 378.36 0.00 2174.10 378.36 3.79 7200.00
eil40 421.10 - - 7200.00 421.10 8.41 7200.00 421.10 8.52 7200.00 421.64 10.72 7200.00
dantzig42 247.25 - - 7200.00 247.25 0.00 676.00 247.25 0.00 999.18 247.25 3.99 7200.00
swiss42 350.45 - - 7200.00 350.45 3.81 7200.00 350.45 3.95 7200.00 350.45 7.62 7200.00
eil45 474.44 - - 7200.00 474.44 13.24 7200.00 476.22 13.27 7200.00 476.22 16.76 7200.00
Avg. 307.12 - - 307.12 1.38 307.20 1.38 307.22 2.49
AvgS 142413 611.40 11787 215.15 11146 273.72 210534 1273.33
Avg. Feas. 292.90 2.60 292.41 0.27 292.41 0.26 292.41 0.87
#Opt. 9 18 18 14
Avg. TGJL16 611.40 18.19 18.55 42.91
Avg. NM B&C 215.15 273.72
Avg. basic Dyn-B&C 84.41 67.94 1273.33

Table 11: Comparison of Integer Solutions with quadratic service times on instances of Set 1 (Taş et al. (2016)).
each method for each instance can be found in the Appendix (available as Supplementary Material
online) for the symmetric and the asymmetric instances. In addition, we report, for each exact
method, the number of instances for which the corresponding method obtained the best known
solution and the number of instances solved to optimality within the time limit.
Moreover, for the large and quadratic service times, we report the results obtained on the
instances of Set 2 (symmetric in Table 14 and asymmetric in Table 17) by the GA+SECs algorithm
by considering 10 runs for each instance. To provide a comparison with the results obtained by
GA+SECs, we also report the upper bound UB and the lower bound LB obtained by the exact
method NMI Dyn-B&C with a time limit of 20000 seconds. For GA+SECs, we show the minimum
and the average values of the solutions obtained over 10 runs, and the average computing time over
the 10 runs.

6.3.1 Symmetric Instances of Set 2

From Table 12 we can observe that the GA algorithm is able to obtain the best solution for all
instances but one, when the best solution over 10 runs is considered. By considering the average
values corresponding to each instance, the average gap is only 0.19%, and the average computing
time 1.26 seconds, thus confirming the effectiveness of the GA algorithm. As it was noted for
the 22 instances of Set 1 considered in Taş et al. (2016), the lower bound obtained at the root
node for the NMI Dyn-B&C algorithm is much larger than that obtained by TGJL16, even though
it requires longer computing times. As regard as the integer solutions found, TGJL16 solves to
optimality within the time limit only 6 instances, while the Dynamic B&C algorithm based on the
NMI formulation solves 12 (out of 13) instances. In addition, the computing time of the latter to
prove the optimality of the solutions is much shorter than that of the former. Analogously, the
average number of nodes explored by NMI Dyn-B&C is significantly smaller than that of TGJL16.
The results reported in Table 13 show that the problem becomes more difficult when we consider
the medium service times. In this case, only 2 instances can be solved to optimality within the time
limit by TGJL16 and 6 by NMI Dyn-B&C. In addition, the computing time on the instances solved
by NMI Dyn-B&C is significantly shorter than that of TGJL16, even when we consider all the
instances solved to optimality within the time limit by each method. With medium service times,
the computing times become larger and for some instances, highlighted by an asterisk, the methods
run out of memory. However, also in the case of medium service times, we can observe that NMI
Dyn-B&C outperforms TGJL16, and that GA+SECs always finds the best known solution. Since
the performances of the exact methods deteriorate with the increase of the service times, we only
report the results obtained by the GA+SECs algorithm with larger and quadratic service times,
and use the upper bound and lower bound values computed by executing NMI Dyn-B&C with
20000 seconds of time limit as a comparison. The results are shown in Table 14. As it can be seen,
in both cases, for all instances, the GA+SECs algorithm always finds the same solution obtained
by NMI Dyn-B&C or a better one. In addition, the computing times are very short, the larger ones
(6 seconds, on average) appearing in the case of quadratic service times.

6.3.2 Asymmetric Instances of Set 2

When we consider the results reported in Table 15 for the asymmetric instances of Set 2 with
small service times, we can note a behavior similar to that observed in Table 12 for the symmetric
instances. The GA algorithm finds several best solutions in short computing time (0.44 seconds on
average). The solution found is proved to be optimal for 26 instances (out of 27) by the Dynamic
B&C algorithm, while TGJL16 can prove the optimality of only 16 solutions. In addition, the

29
 GA+SECs TGJL16 NMI Dyn-B&C
# inst Best Val Avg% Min% Time LB Time UB Gap% Time LB Time UB Gap% Time
att35 271.36 271.36 0.00 0.00 0.74 238.07 0.02 271.36 0.00 76.89 264.96 0.73 271.36 0.00 3.05
dantzig35 296.02 296.02 0.14 0.00 0.66 242.80 0.03 296.02 0.00 419.88 285.84 0.56 296.02 0.00 10.42
hk35 347.51 347.51 0.00 0.00 0.64 303.97 0.04 347.51 0.00 112.29 337.31 0.65 347.51 0.00 6.13
att40 296.33 296.33 0.00 0.00 0.90 247.62 0.04 296.33 0.00 5293.41 288.89 0.88 296.33 0.00 9.26
hk40 356.28 356.28 0.00 0.00 0.93 304.76 0.04 356.28 0.00 1404.32 343.39 0.90 356.28 0.00 31.04
att45 318.94 318.94 0.06 0.00 1.24 258.23 0.04 318.94 2.09 50000.00 309.84 0.95 318.94 0.00 473.42
hk45 365.41 365.41 0.00 0.00 1.27 301.77 0.04 365.41 0.00 30916.40 353.13 1.63 365.41 0.00 487.68
att48 336.75 336.75 0.89 0.00 1.42 271.45 0.04 338.97 5.58 50000.00 326.51 1.93 336.75 0.00 3311.93
gr48 383.13 383.13 0.62 0.00 1.44 311.41 0.05 383.13 3.70 50000.00 362.01 2.40 383.13 0.00 13732.86

30
hk48 371.86 371.86 0.01 0.00 1.57 320.20 0.05 371.86 1.25 50000.00 358.56 1.93 371.86 0.00 3621.10
eil51 442.53 443.28 0.70 0.17 1.70 392.49 0.17 442.53 0.25 50000.00 429.21 4.25 442.53 0.00 24122.59
berlin52 439.76 439.76 0.00 0.00 1.95 356.02 0.03 439.76 6.31 50000.00 416.05 2.05 439.76 1.30 50000.00
brazil58 386.48 386.48 0.05 0.00 1.95 273.53 0.07 386.48 5.18 50000.00 369.57 2.13 386.48 0.00 43579.73
Avg. 354.80 354.85 0.19 0.01 1.26 294.02 0.05 354.97 1.87 341.94 1.62 354.80 0.10
AvgS 1726169 6370.53 228949 7449.10
#Best 6 12 12 13
#Opt. 6 12
Avg. TGJL16 6370.53 91.26

Table 12: Comparison of Lower Bounds and Integer Solutions with small service times on symmetric instances of Set 2.
 GA+SECs TGJL16 NMI Dyn-B&C
# inst Best Val Avg% Min% Time LB UB Gap% Time LB UB Gap% Time
att35 294.16 294.16 0.00 0.00 0.73 268.07 294.16 0.00 12929.82 280.38 294.16 0.00 57.63
dantzig35 323.05 323.05 0.80 0.00 0.59 293.19 323.05 3.33 50000.00 302.22 323.05 0.00 1753.73
hk35 376.25 376.25 0.00 0.00 0.64 347.21 376.25 0.00 26673.53 357.23 376.25 0.00 236.61
att40 324.04 324.04 0.00 0.00 0.72 288.42 324.04 4.50 50000.00 308.08 324.04 0.00 1034.17
hk40 393.46 393.46 0.00 0.00 0.92 362.47 393.46 4.35 50000.00 368.98 393.46 0.00 18520.97
att45 351.93 351.93 0.00a 0.00 0.99 314.28 351.93 6.04 50000.00 333.94 351.93 1.58 50000.00
hk45 409.73 409.73 0.00 0.00 1.27 376.21 409.73 5.79 50000.00 382.58 409.73 3.26 50000.00
att48 374.30 374.30 1.32 0.00 1.37 333.04 374.30 6.17 50000.00 353.82 374.30 2.50 50000.00
gr48 428.64 428.64 1.22 0.00 1.25 376.67 431.09 8.06 50000.00 396.31 428.64 0.00 21826.02
hk48 420.36 420.36 0.00 0.00 0.89 380.58 420.36 5.93 50000.00 392.38 420.36 4.17 50000.00

31
eil51 503.35 503.35 0.52 0.00 1.68 472.71 503.35 3.82 23540.00* 480.24 503.35 3.14 43095.26*
berlin52 497.94 497.94 0.00 0.00 1.75 427.80 497.94 12.66 50000.00 444.91 497.94 7.88 50000.00
brazil58 429.80 429.80 0.00a 0.00 1.93 342.63 429.80 10.30 15876.20* 392.25 429.80 5.97 50000.00
Avg. 394.39 394.39 0.30 0.00 1.13 352.56 394.57 5.46 368.72 394.39 2.19
AvgS 8499711 19801.67 278932 7238.19
#Best 9 13 12 13
#Opt. 2 6
Avg. TGJL16 19801.67 147.12

Table 13: Comparison of Lower Bounds and Integer Solutions with medium service times on symmetric instances of Set 2.
a
the gap is ≥ 0.001% and ≤ 0.005%
 Large service times Quadratic service times
NMI Dyn-B&C GA+SECs NMI Dyn-B&C GA+SECs
# inst UB LB Val ValAvg Time UB LB Val ValAvg Time
att35 347.84 341.06 347.84 347.84 0.58 263.18 263.18 263.18 263.18 2.66
dantzig35 386.92 355.76 386.92 388.23 0.53 289.40 289.40 289.40 291.01 3.66
hk35 444.91 425.54 444.91 448.09 0.53 341.36 341.36 341.36 341.36 2.32
att40 390.04 364.66 390.04 390.04 0.71 288.41 288.41 288.41 288.41 4.02
hk40 480.81 442.14 480.81 481.69 0.58 351.77 351.77 351.77 351.77 3.60
att45 431.72 399.31 431.72 431.97 0.98 312.30 307.35 312.30 312.54 7.23
hk45 515.09 465.11 515.09 517.65 0.91 363.02 355.98 363.02 363.02 4.80

32
att48 468.25 427.94 468.25 472.54 1.22 331.12 317.43 331.12 335.86 8.05
gr48 534.64 483.34 534.64 538.57 1.13 393.97 350.82 390.02 392.84 6.54
hk48 533.07 483.13 533.07 536.96 0.95 370.84 352.38 370.84 371.11 8.26
eil51 656.49 610.08 656.49 663.62 1.60 471.88 400.66 471.72 473.73 7.42
berlin52 613.14 527.38 613.03 617.11 0.87 460.40 410.05 460.40 460.40 6.29
brazil58 539.07 458.59 539.07 539.09 1.96 390.56 373.50 390.56 390.56 13.30
Avg. 487.84 444.92 487.84 490.26 0.97 356.02 338.64 355.70 356.60 6.01
#Best 12 13 3 11 13 7

Table 14: GA with large and quadratic service times on symmetric instances of Set 2.
computing time to prove the optimality of the solutions solved by TGJL16 is much shorter for
the NMI Dyn-B&C algorithm (on average about 37 seconds versus 4935 seconds). Similarly, the
average number of nodes explored by NMI Dyn-B&C is significantly smaller than that of TGJL16.
Therefore, we can conclude that, with small service times, the proposed algorithm turns out to be
very effective on the asymmetric instances as well.
From Table 16 we can see that, with medium service times, the performance of NMI Dyn-B&C
is still much better than that of TGJL16, since the latter can solve to optimality within the time
limit only 5 instances compared to 16 found by the former. In addition, the computing times of NMI
Dyn-B&C are much smaller than those of TGJL16. Similar to what happens for the symmetric
instances, the medium service time case is harder than the small service time one, and, for some
instances, the methods run out of memory. We also observe that GA+SECs shows a very good
performance, being able to obtain 18 best known solutions (over 27 instances) with an average gap
of 0.16% in very short computing times.
Similar to what we did for the symmetric instances of Set 2, we also report the performance
of the GA+SECs algorithm on the asymmetric instances of Set 2 with large and quadratic service
times, and compare these results with those obtained by NMI Dyn-B&C with a time limit of 20000
seconds. As it can be seen from Table 17, although GA+SECs obtains a smaller number of best
solution values w.r.t. NMI Dyn-B&C, the difference between the solution values is always very
small, and the computing times are consistently short. More precisely, the average gap between
the GA+SECs and the NMI Dyn-B&C solution values is 0.10% for the large and 0.11% for the
quadratic service times.

6.4 Genetic Algorithm on instances of Set 3

In this section, we consider the instances of Set 3 and briefly comment, in Section 6.4.1, on the
results obtained by the GA+SECs algorithm on the symmetric instances, and, in Section 6.4.2,
the corresponding results for the asymmetric instances. The detailed results are reported in the
Appendix (available as Supplementary Material online) both for the symmetric and asymmetric
instances with small, medium, large and quadratic service times. Instances of Set 3 are the largest
ones, hence we did not execute the exact methods on them, and to evaluate the results obtained
by GA+SECs we report the lower bound values obtained by NMI Dyn-B&C at the root node and
the results obtained by the randomized NNH algorithm described in Section 5.1. Both GA+SECs
and NNH are executed 10 times for each instance. In each table, we report the instance name, and,
for the corresponding service time, the lower bound value, the best solution value found by NNH
out of 10 runs, and, for GA+SECs, the average and minimum solution values obtained over the
10 runs, and the average computing time (expressed in seconds) over the 10 runs. The computing
time of NNH is not reported, since it is always negligible.

6.4.1 Symmetric Instances of Set 3

The results obtained for small and medium service times show that significant improvements can be
found by applying GA+SECs with respect to NNH (the average minimum solution values are 468.31
versus 549.67 for the small service times, and 530.32 versus 615.04 for the medium ones), and the
computing times are about 35 seconds on average. We can observe that the average solution values
found by GA+SECs are not very different from the minimum ones, showing that the algorithm is
rather stable. We also note that the average percentage gaps between the solution values found
by GA+SECs and the lower bound values are rather large: this is not surprising since a similar
behavior happens for the instances of Set 2. However, we also observe that, in the latter case,

33
 GA+SECs TGJL16 NMI Dyn-B&C
# inst Best Val Avg% Min% Time LB Time UB Gap% Time LB Time UB Gap% Time
br17 88.60 88.60 0.00 0.00 0.16 27.01 0.00 88.60 0.00 2.00 87.18 0.09 88.60 0.00 1.53
ft30 338.54 338.54 0.40 0.00 0.23 278.97 0.02 338.54 0.00 1005.68 323.87 0.15 338.54 0.00 14.12
ftv30a 298.60 298.60 0.14 0.00 0.28 268.60 0.02 298.60 0.00 11.03 289.10 0.11 298.60 0.00 1.12
ftv30b 326.36 326.36 0.34 0.00 0.34 303.34 0.02 326.36 0.00 12.95 316.91 0.14 326.36 0.00 1.05
ftv30c 305.95 305.95 0.00 0.00 0.30 268.66 0.02 305.95 0.00 101.96 291.18 0.22 305.95 0.00 1.96
p30 201.96 201.96 0.01 0.00 0.27 144.48 0.01 201.96 2.51 50000.00 196.18 0.21 201.96 0.00 46.82
ry30p 303.02 303.02 0.46 0.00 0.33 247.13 0.02 303.02 0.00 993.95 291.65 0.18 303.02 0.00 11.04
ftv33 328.01 328.01 0.00 0.00 0.44 292.87 0.02 328.01 0.00 21.88 319.75 0.33 328.01 0.00 1.92
ft35 394.62 394.62 1.55 0.00 0.30 337.02 0.02 394.62 0.00 46230.96 376.45 0.25 394.62 0.00 270.68
ftv35 358.07 360.08 1.41 0.56 0.43 322.67 0.03 358.07 0.00 2452.24 342.86 0.42 358.07 0.00 35.33
ftv35a 341.12 341.12 0.94 0.00 0.40 306.64 0.02 341.12 0.00 384.90 325.36 0.29 341.12 0.00 11.78
ftv35b 351.27 356.04 1.61 1.36 0.42 320.41 0.03 351.27 0.00 647.08 335.42 0.37 351.27 0.00 7.50
ftv35c 343.30 343.30 0.00 0.00 0.52 308.72 0.03 343.30 0.00 667.67 330.51 0.38 343.30 0.00 16.12
p35 229.24 229.24 0.38 0.00 0.39 160.62 0.01 238.05 9.19 50000.00 222.10 0.39 229.24 0.00 46.63
ry35p 326.29 326.29 0.00 0.00 0.43 297.81 0.03 326.29 0.00 108.50 320.52 0.37 326.29 0.00 2.28
ftv38 380.31 381.83 0.84 0.40 0.51 342.64 0.04 380.31 0.00 20789.05 363.64 0.44 380.31 0.00 127.61

34
ft40 436.93 436.93 0.76 0.00 0.36 374.07 0.03 436.93 2.82 50000.00 416.86 0.45 436.93 0.00 3593.72
ftv40a 365.59 365.59 0.00 0.00 0.54 329.82 0.03 365.59 0.00 2599.55 350.24 0.60 365.59 0.00 53.58
ftv40b 372.98 372.98 0.00 0.00 0.54 341.56 0.04 372.98 0.00 2939.59 358.27 0.48 372.98 0.00 40.31
ftv40c 377.38 379.36 1.22 0.53 0.53 337.17 0.03 377.38 1.95 50000.00 358.50 0.45 377.38 0.00 953.51
ry40p 365.59 365.68 0.07 0.02 0.61 320.17 0.03 365.59 0.77 50000.00 352.77 0.68 365.59 0.00 58.44
p43 288.45 288.45 0.16 0.00 0.44 43.40 0.02 288.45 0.67 50000.00 287.52 0.50 288.45 0.22 50000.00
ftv44 395.91 398.14 1.99 0.56 0.53 354.54 0.06 395.91 3.25 50000.00 370.78 0.45 395.91 0.00 8936.01
ft45 473.37 473.37 1.30 0.00 0.45 402.27 0.04 473.37 3.88 50000.00 450.34 0.58 473.37 0.00 7537.42
ftv45b 408.17 410.63 2.55 0.60 0.55 359.25 0.06 408.17 3.65 50000.00 385.63 1.17 408.17 0.00 30052.82
ftv45c 371.13 374.31 0.88 0.86 0.91 326.22 0.05 371.13 3.20 50000.00 351.23 0.90 371.13 0.00 2081.32
ry45p 394.23 394.23 1.37 0.00 0.68 338.33 0.03 394.23 2.34 50000.00 379.70 0.93 394.23 0.00 383.69
Avg. 339.44 340.12 0.68 0.18 0.44 287.20 0.03 339.77 1.27 325.72 0.43 339.44 0.01
AvgS 1480765 4935.56 55680 2088.01
#Best 7 19 26 27
#Opt. 16 26
Avg. TGJL16 4935.56 37.37

Table 15: Comparison of Lower Bounds and Integer Solutions with small service times on asymmetric instances of Set 2.
 GA+SECs TGJL16 NMI Dyn-B&C
# inst Best Val Avg% Min% Time LB UB Gap% Time LB UB Gap% Time
br17 92.09 92.09 0.00 0.00 0.19 68.02 92.09 0.00 2.05 89.11 92.09 0.00 2.71
ft30 364.80 364.80 0.51 0.00 0.23 309.24 364.80 4.12 50000.00 337.24 364.80 0.00 1463.17
ftv30a 321.26 321.26 0.37 0.00 0.27 294.25 321.26 0.00 330.24 304.81 321.26 0.00 12.46
ftv30b 349.82 349.82 0.00 0.00 0.33 322.75 349.82 0.00 748.19 334.28 349.82 0.00 15.59
ftv30c 331.12 331.12 0.00 0.00 0.30 296.54 331.12 0.00 25885.07 309.21 331.12 0.00 22.07
p30 212.71 212.71 0.00 0.00 0.27 161.65 212.71 9.56 50000.00 204.35 212.71 0.00 836.31
ry30p 325.09 325.09 0.14 0.00 0.30 285.80 325.09 3.38 50000.00 307.15 325.09 0.00 412.95
ftv33 355.49 355.49 0.00 0.00 0.45 315.30 355.49 0.00 4291.36 337.60 355.49 0.00 23.07
ft35 426.24 426.24 1.58 0.00 0.30 365.34 426.24 6.79 50000.00 394.74 426.24 2.07 50000.00
ftv35 390.43 392.68 1.18 0.57 0.42 345.99 390.43 6.09 50000.00 362.59 390.43 0.00 11480.23
ftv35a 372.28 372.28 1.09 0.00 0.39 332.29 372.28 4.02 50000.00 346.14 372.28 0.00 2017.69
ftv35b 380.78 380.78 1.51 0.00 0.41 342.43 380.78 2.89 50000.00 356.63 380.78 0.00 501.17
ftv35c 373.66 373.66 0.02 0.00 0.37 335.71 373.66 4.60 50000.00 349.88 373.66 0.00 3474.43
p35 243.14 243.14 0.38 0.00 0.36 175.54 244.37 11.26 50000.00 231.71 243.14 0.00 1009.46
ry35p 354.38 354.38 0.00 0.00 0.43 326.71 354.38 0.48 50000.00 340.97 354.38 0.00 29.06
ftv38 417.91 420.90 1.30 0.72 0.50 369.23 417.91 7.56 50000.00 386.74 417.91 1.97 50000.00

35
ft40 478.38 478.45 0.53 0.01 0.36 408.64 478.38 8.37 50000.00 438.76 478.38 4.45 50000.00
ftv40a 405.49 405.49 0.03 0.00 0.54 362.27 405.49 6.09 50000.00 374.53 405.49 1.47 50000.00
ftv40b 412.11 412.11 0.00 0.00 0.54 369.78 412.11 6.71 50000.00 380.35 412.11 1.31 50000.00
ftv40c 411.78 415.18 2.42 0.83 0.50 364.51 411.78 7.23 34262.71* 383.44 411.78 0.00 756.48
ry40p 401.18 401.27 0.21 0.02 0.60 359.92 401.18 5.69 50000.00 380.49 401.18 0.00 12146.50
p43 291.96 291.96 0.15 0.00 0.43 47.13 292.21 1.34 50000.00 289.87 291.96 0.46 47157.33*
ftv44 446.13 448.35 1.07 0.50 0.55 392.21 446.13 10.53 7821.86* 401.06 446.13 6.30 50000.00
ft45 526.36 527.04 1.80 0.13 0.46 443.93 526.36 9.95 50000.00 478.29 526.36 5.93 50000.00
ftv45b 456.53 458.77 2.37 0.49 0.54 402.23 456.53 9.19 50000.00 415.46 456.85 6.41 50000.00
ftv45c 414.52 418.88 1.45 1.05 0.89 363.62 414.52 9.38 50000.00 377.96 414.52 4.60 50000.00
ry45p 437.30 437.30 1.28 0.00 0.69 391.10 437.30 6.67 50000.00 413.71 437.30 1.23 50000.00
Avg. 370.11 370.79 0.72 0.16 0.43 316.75 370.16 5.26 345.45 370.12 1.34
AvgS 2582474 6251.38 113910 2137.71
#Best 7 18 25 26
#Opt. 5 16
Avg. TGJL16 6251.38 15.18

Table 16: Comparison of Lower Bounds and Integer Solutions with medium service times on asymmetric instances of Set 2.
 Large service times Quadratic service times
NMI Dyn-B&C GA+SECs NMI Dyn-B&C GA+SECs
# inst UB LB Val ValAvg Time UB LB Val ValAvg Time
br17 99.61 99.61 99.61 99.61 0.15 85.63 85.63 85.63 85.63 0.37
ft30 422.00 391.09 422.00 427.64 0.22 337.08 337.08 337.08 339.36 1.16
ftv30a 372.55 372.55 373.32 374.33 0.28 294.44 294.44 294.44 294.60 1.14
ftv30b 403.72 403.72 403.72 403.72 0.33 323.21 323.21 323.21 323.21 1.21
ftv30c 389.61 375.29 389.61 389.75 0.30 300.92 300.92 300.92 301.00 1.08
p30 237.70 237.70 237.70 237.73 0.27 194.07 192.56 194.07 194.43 1.51
ry30p 375.81 361.50 375.81 375.98 0.30 298.30 298.30 298.68 300.67 2.01
ftv33 419.49 407.90 419.49 419.49 0.45 327.11 327.11 327.11 327.11 2.00
ft35 500.69 451.01 504.28 512.05 0.30 395.85 385.90 396.25 400.18 1.43
ftv35 465.54 427.42 465.54 470.18 0.41 361.73 358.32 362.26 365.54 2.18
ftv35a 443.16 411.13 443.90 448.50 0.37 341.50 341.50 341.50 349.65 1.62
ftv35b 450.30 423.13 450.30 458.28 0.42 349.74 349.74 349.74 355.53 1.85
ftv35c 440.66 413.31 440.66 442.88 0.36 340.49 340.49 340.49 340.49 1.94
p35 276.50 270.44 276.50 276.99 0.38 219.50 218.72 219.50 219.55 2.94

36
ry35p 420.66 417.86 420.66 420.66 0.43 322.79 322.79 322.79 322.79 2.48
ftv38 507.81 458.48 508.51 515.25 0.48 387.19 367.29 387.19 389.86 2.35
ft40 576.25 508.96 577.28 579.77 0.36 444.83 422.55 444.84 447.40 1.66
ftv40a 501.68 495.81 501.68 501.94 0.54 371.08 354.31 371.08 371.23 2.45
ftv40b 505.97 449.81 507.02 507.02 0.54 377.82 368.25 377.82 377.82 3.31
ftv40c 494.86 450.99 494.86 513.54 0.44 376.60 363.27 381.93 386.28 1.75
ry40p 487.20 458.74 487.20 487.86 0.60 366.14 357.67 366.14 367.10 3.62
p43 300.76 296.25 300.96 301.50 0.44 288.17 285.52 288.18 288.29 4.83
ftv44 567.22 485.93 567.22 573.68 0.53 410.78 368.83 411.66 417.90 2.27
ft45 651.52 562.35 658.88 665.61 0.45 491.37 449.61 493.35 501.59 2.83
ftv45b 572.00 495.22 573.02 580.95 0.55 419.45 381.99 422.38 432.64 3.48
ftv45c 516.41 455.39 515.37 527.23 0.80 374.43 349.92 374.43 376.68 3.26
ry45p 544.02 506.58 544.02 550.95 0.67 399.62 375.40 399.53 405.49 6.08
Avg. 442.36 410.67 442.93 446.78 0.42 340.73 330.42 341.19 343.78 2.33
#Best 26 18 4 26 18 6

Table 17: GA with large and quadratic service times on asymmetric instances of Set 2.
the solution values found by GA+SECs are often close to the best known or even optimal ones.
Similar considerations can be done for the large and quadratic service times, although the latter
case requires one order of magnitude larger computing times.

6.4.2 Asymmetric Instances of Set 3

As shown in the Appendix, also for these instances, GA+SECs obtains, in short computing times,
significant improvements over the randomized NNH for all the instances. The largest computing
time is required by the case of quadratic service times, the average computing time being about 50
seconds, while for all the other cases it is less than 5 seconds.

7 Conclusions and Future Research

We studied the Traveling Salesman Problem with Time-dependent Service times (TSP-TS), which
considers the service time at each customer as a continuous function of the start time of service. We
proposed a new formulation for the problem and included explicit subtour elimination constraints,
dynamically separated. In addition, we proposed an upper bound on the total route duration,
obtained by a multi-operator Genetic Algorithm, an improved lower bound on the total service time,
and new lower and upper bounds on the start time of service at each customer. These ingredients
are included in two Branch-and-Cut algorithms, one of which exploits the dynamic update of the
bounds during the solving process. The proposed algorithms are tested on benchmark instances
from the literature and compared to an existing method. The results show that the optimality
of the solutions found can be proved for a larger set of instances in shorter computing times.
Additional computational experiments of the proposed exact algorithm have been conducted on
larger size symmetric instances with up to 58 nodes and on asymmetric instances with up to 45
nodes showing its effectiveness. Finally, the genetic algorithm has been tested on symmetric and
asymmetric instances with up to 200 nodes, obtaining good quality solutions in short computing
times.
Future research will focus on extending the proposed methods to other variants of the TSP that
embed the time-dependency. In addition, the problem with time-dependent service times could be
generalized to deal with other features, such as more vehicles or time window constraints.

Acknowledgments
This material is based upon work supported by the Air Force Office of Scientific Research under
award number FA9550-17-1-0025. We also acknowledge project CONICYT PFCHA/DOCTORADO
BECAS CHILE/2015-72160389. We are very grateful to the reviewers for their insightful comments
and, in particular, to one of them for showing the connection between TSP-TS and TDTSP.

37
References
Abeledo, H., Fukasawa, R., Pessoa, A., and Uchoa, E. (2013). The time dependent traveling
salesman problem: polyhedra and algorithm. Mathematical Programming Computation, 5(1):27–
55.

Albayrak, M. and Allahverdi, N. (2011). Development a new mutation operator to solve the
Traveling Salesman Problem by aid of Genetic Algorithms. Expert Systems with Applications,
38(3):1313–1320.

Arigliano, A., Calogiuri, T., Ghiani, G., and Guerriero, E. (2018). A branch-and-bound algorithm
for the time-dependent travelling salesman problem. Networks, 72:382–392.

Arigliano, A., Ghiani, G., Grieco, A., Guerriero, E., and Plana, I. (2019). Time-dependent asymmet-
ric traveling salesman problem with time windows: Properties and an exact algorithm. Discrete
Applied Mathematics, 261:28–39.

Banzhaf, W. (1990). The “molecular” traveling salesman. Biological Cybernetics, 64(1):7–14.

Bigras, L.-P., Gamache, M., and Savard, G. (2008). The time-dependent traveling salesman prob-
lem and single machine scheduling problems with sequence dependent setup times. Discrete
Optimization, 5(4):685–699.

Boland, N., Hewitt, M., Marshall, L., and Savelsbergh, M. (2017). The continuous-time service
network design problem. Operations Research, 65(5):1303–1321.

Buriol, L., França, P., and Moscato, P. (2004). A new memetic algorithm for the asymmetric
traveling salesman problem. Journal of Heuristics, 10(5):483–506.

Contreras-Bolton, C. and Parada, V. (2015). Automatic combination of operators in a genetic

algorithm to solve the traveling salesman problem. PloS one, 10(9):e0137724.

Cordeau, J.-F., Ghiani, G., and Guerriero, E. (2012). Analysis and branch-and-cut algorithm for
the time-dependent travelling salesman problem. Transportation Science, 48(1):46–58.

Dantzig, G., Fulkerson, R., and Johnson, S. (1954). Solution of a large-scale traveling-salesman
problem. Journal of the Operations Research Society of America, 2(4):393–410.

Desrochers, M. and Laporte, G. (1991). Improvements and extensions to the miller-tucker-zemlin

subtour elimination constraints. Operations Research Letters, 10(1):27–36.

Fogel, D. (1988). An evolutionary approach to the traveling salesman problem. Biological Cyber-
netics, 60(2):139–144.

Fogel, D. (1993). Applying evolutionary programming to selected traveling salesman problems.

Cybernetics and Systems, 24(1):27–36.

Gavish, B. and Graves, S. (1978). The travelling salesman problem and related problems. Technical
Report GR-078-78, Operations Research Center, MIT.

Gouveia, L. and Voß, S. (1995). A classification of formulations for the (time-dependent) traveling
salesman problem. European Journal of Operational Research, 83(1):69–82.

38
Grefenstette, J., Gopal, R., Rosmaita, B., and Gucht, D. (1985). Genetic algorithms for the travel-
ing salesman problem. In Proceedings of the 1st International Conference on Genetic Algorithms,
pages 160–168, Hillsdale, NJ, USA. L. Erlbaum Associates Inc.

Ichoua, S., Gendreau, M., and Potvin, J.-Y. (2003). Vehicle dispatching with time-dependent travel
times. European Journal of Operational Research, 144(2):379–396.

Kinable, J., Cire, A. A., and van Hoeve, W.-J. (2017). Hybrid optimization methods for time-
dependent sequencing problems. European Journal of Operational Research, 259(3):887–897.

Maffioli, F. and Sciomachen, A. (1997). A mixed-integer model for solving ordering problems with
side constraints. Annals of Operations Research, 69:277–297.

Miller, C., Tucker, A., and Zemlin, R. (1960). Integer programming formulation of traveling sales-
man problems. Journal of the ACM, 7(4):326–329.

Miranda-Bront, J., Méndez-Dı́az, I., and Zabala, P. (2014). Facets and valid inequalities for
the time-dependent travelling salesman problem. European Journal of Operational Research,
236(3):891–902.

Montero, A., Méndez-Dı́az, I., and Miranda-Bront, J. J. (2017). An integer programming approach
for the time-dependent traveling salesman problem with time windows. Computers & Operations
Research, 88:280–289.

Mühlenbein, H. (1991). Parallel genetic algorithms, population genetics and combinatorial opti-
mization. In Becker, J., Eisele, I., and Mündemann, F., editors, Parallelism, Learning, Evolution,
volume 565 of Lecture Notes in Computer Science, pages 398–406. Springer Berlin Heidelberg.

Öncan, T., Altınel, İ. K., and Laporte, G. (2009). A comparative analysis of several asymmetric
traveling salesman problem formulations. Computers & Operations Research, 36(3):637–654.

Padberg, M. and Rinaldi, G. (1990). An efficient algorithm for the minimum capacity cut problem.
Mathematical Programming, 47(1-3):19–36.

Picard, J.-C. and Queyranne, M. (1978). The time-dependent traveling salesman problem and its
application to the tardiness problem in one-machine scheduling. Operations Research, 26(1):86–
110.

Reinelt, G. (1991). Tsplib-a traveling salesman problem library. ORSA Journal on Computing,
3(4):376–384.

Reisleben, B., Merz, P., and Freisleben, B. (1996). A genetic local search algorithm for solving
symmetric and asymmetric traveling salesman problems. In Proceedings of IEEE International
Conference on Evolutionary Computation, pages 616–621.

Roberti, R. and Toth, P. (2012). Models and algorithms for the asymmetric traveling salesman
problem: an experimental comparison. EURO Journal on Transportation and Logistics, 1(1-
2):113–133.

Sun, P., Veelenturf, L. P., Dabia, S., and Van Woensel, T. (2018). The time-dependent capacitated
profitable tour problem with time windows and precedence constraints. European Journal of
Operational Research, 264(3):1058–1073.

39
 Syswerda, G. (1991). Schedule optimization using genetic algorithms. Handbook of Genetic Algo-
rithms, pages 332–349.

Taş, D., Gendreau, M., Jabali, O., and Laporte, G. (2016). The traveling salesman problem with
time-dependent service times. European Journal of Operational Research, 248(2):372–383.

Vander Wiel, R. J. and Sahinidis, N. (1996). An exact solution approach for the time-dependent
traveling-salesman problem. Naval Research Logistics (NRL), 43(6):797–820.

Vu, D., Hewitt, M., Boland, N., and Savelsbergh, M. (2018). Solving time dependent traveling
salesman problems with time windows. Technical report, Optimization Online.

Wang, Y., Sun, J., Li, J., and Gao, K. (2012). A Modified Inver-over Operator for the Traveling
Salesman Problem. In Huang, D.-S., Gan, Y., Gupta, P., and Gromiha, M., editors, Advanced
Intelligent Computing Theories and Applications. With Aspects of Artificial Intelligence, volume
6839 of Lecture Notes in Computer Science, pages 17–23. Springer Berlin Heidelberg.

40

View publication stats